Laser Scanning Systems in Highway and Safety Assessment: Analysis of Highway Geometry and Safety Using LiDAR [1st ed.] 978-3-030-10373-6;978-3-030-10374-3

Table of contents :
Front Matter ....Pages i-xv
Front Matter ....Pages 1-1
Laser Scanning Technologies in Road Geometry Modeling (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 3-13
Road Geometric Modeling Using Laser Scanning Data: A Critical Review (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 15-31
Road Geometric Modeling Using a Novel Hierarchical Approach (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 33-46
Optimizing Support Vector Machine and Ensemble Trees Using Taguchi Method for Road Extraction from LiDAR Data (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 47-60
An Integrated Machine Learning Approach for Automatic Highway Extraction from Airborne LiDAR Data and Orthophotos (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 61-76
Effect of Roadside Features on Injury Severity of Traffic Accidents (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 77-86
Novel GIS-Based Model for Automatic Identification of Road Geometry in Vector Data (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 87-94
Front Matter ....Pages 95-95
Review of Traffic Accident Predictions with Neural Networks (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 97-109
Modeling Traffic Accident Severity Using Neural Networks and Support Vector Machines (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 111-117
Predicting Injury Severity of Road Traffic Accidents Using a Hybrid Extreme Gradient Boosting and Deep Neural Network Approach (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 119-127
Applications of Deep Learning in Severity Prediction of Traffic Accidents (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 129-139
Forecasting Severity of Motorcycle Crashes Using Transfer Learning (Biswajeet Pradhan, Maher Ibrahim Sameen)....Pages 141-157

Citation preview

Advances in Science, Technology & Innovation IEREK Interdisciplinary Series for Sustainable Development

Biswajeet Pradhan Maher Ibrahim Sameen

Laser Scanning Systems in Highway and Safety Assessment Analysis of Highway Geometry and Safety Using LiDAR

Advances in Science, Technology & Innovation IEREK Interdisciplinary Series for Sustainable Development Editorial Board Members Anna Laura Pisello, Department of Engineering, University of Perugia, Italy Dean Hawkes, Cardiff University, UK Hocine Bougdah, University for the Creative Arts, Farnham, UK Federica Rosso, Sapienza University of Rome, Rome, Italy Hassan Abdalla, University of East London, London, UK Sofia-Natalia Boemi, Aristotle University of Thessaloniki, Greece Nabil Mohareb, Beirut Arab University, Beirut, Lebanon Saleh Mesbah Elkaffas, Arab Academy for Science, Technology, Egypt Emmanuel Bozonnet, University of la Rochelle, La Rochelle, France Gloria Pignatta, University of Perugia, Italy Yasser Mahgoub, Qatar University, Qatar Luciano De Bonis, University of Molise, Italy Stella Kostopoulou, Regional and Tourism Development, University of Thessaloniki, Thessaloniki, Greece Biswajeet Pradhan, Faculty of Engineering and IT, University of Technology Sydney, Sydney, Australia Md. Abdul Mannan, Universiti Malaysia Sarawak, Malaysia Chaham Alalouch, Sultan Qaboos University, Muscat, Oman Iman O. Gawad, Helwan University, Egypt Series Editor Mourad Amer, Enrichment and Knowledge Exchange, International Experts for Research, Cairo, Egypt

Advances in Science, Technology & Innovation (ASTI) is a series of peer-reviewed books based on the best studies on emerging research that redefines existing disciplinary boundaries in science, technology and innovation (STI) in order to develop integrated concepts for sustainable development. The series is mainly based on the best research papers from various IEREK and other international conferences, and is intended to promote the creation and development of viable solutions for a sustainable future and a positive societal transformation with the help of integrated and innovative science-based approaches. Offering interdisciplinary coverage, the series presents innovative approaches and highlights how they can best support both the economic and sustainable development for the welfare of all societies. In particular, the series includes conceptual and empirical contributions from different interrelated fields of science, technology and innovation that focus on providing practical solutions to ensure food, water and energy security. It also presents new case studies offering concrete examples of how to resolve sustainable urbanization and environmental issues. The series is addressed to professionals in research and teaching, consultancies and industry, and government and international organizations. Published in collaboration with IEREK, the ASTI series will acquaint readers with essential new studies in STI for sustainable development.

More information about this series at http://www.springer.com/series/15883

Biswajeet Pradhan
Maher Ibrahim Sameen

Laser Scanning Systems in Highway and Safety Assessment Analysis of Highway Geometry and Safety Using LiDAR

123

Biswajeet Pradhan Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS) University of Technology Sydney Sydney, Australia

Maher Ibrahim Sameen Centre for Advanced Modelling and Geospatial Information Systems (CAMGIS) University of Technology Sydney Sydney, Australia

ISSN 2522-8714 ISSN 2522-8722 (electronic) Advances in Science, Technology & Innovation IEREK Interdisciplinary Series for Sustainable Development ISBN 978-3-030-10373-6 ISBN 978-3-030-10374-3 (eBook) https://doi.org/10.1007/978-3-030-10374-3 Library of Congress Control Number: 2019933705 © Springer Nature Switzerland AG 2020 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

Recent developments in laser scanning technologies have provided innovative solutions for acquiring three-dimensional (3D) point clouds about road corridors and its environments. Unlike traditional field surveying, satellite imagery, and aerial photography, laser scanning systems offer unique solutions for collecting dense point clouds with millimeter accuracy and in a reasonable time. The data acquired by laser scanning systems empower modeling road geometry and delineating road design parameters such as slope, superelevation, and vertical and horizontal alignments. These geometric parameters have several geospatial applications such as road safety management. The purpose of this book is to promote the core understanding of suitable geospatial tools and techniques for modeling of road traffic accidents by the state-of-the-art artificial intelligence (AI) approaches such as neural networks (NNs) and deep learning (DL) using traffic information and road geometry delineated from laser scanning data. Data collection and management in databases play a major role in modeling and developing predictive tools. Therefore, the first two chapters of this book introduce laser scanning technology with creative explanation and graphical illustrations and review the recent methods of extracting geometric road parameters. The third and fourth chapters present an optimization of support vector machine and ensemble tree methods as well as novel hierarchical object-based methods for extracting road geometry from laser scanning point clouds. Information about historical traffic accidents and their circumstances, traffic (volume, type of vehicles), road features (grade, superelevation, curve radius, lane width, speed limit, etc.) pertains to what is observed to exist on road segments or road intersections. Soft computing models such as neural networks are advanced modeling methods that can be related to traffic and road features to the historical accidents and generates regression equations that can be used in various phases of road safety management cycle. The regression equations produced by NN can identify unsafe road segments, estimate how much safety has changed following a change in design, and quantify the effects of road geometric features and traffic information on road safety. This book aims to help graduate students, professionals, decision makers, and road planners in developing better traffic accident prediction models using advanced neural networks. This book is organized into twelve chapters. Chapter 1 presents an overview of LiDAR technology for geometric road modeling. The concepts of airborne, mobile, and terrestrial systems were explained. The strengths and weaknesses of these systems were discussed in general and in the context of road modeling. The detail literature review indicates that the three LiDAR systems (ALS, MLS, and TLS) share similarity in hardware components (GNSS, IMU, and laser scanner), range measurements principles, and supporting digital cameras. In contrast, the literature showed that these systems differ in accuracy, the safety of data collection, and overall efficiency. Regarding road geometry modeling, the ground-based systems are easier to operate than ALS because the latter requires integration of IMU and GNSS sensors for accurate point georeferencing. Ground-based systems offer more options for setup locations including away from the road. Additionally, MLS and TLS systems provide significantly improved horizontal accuracy due to looking angle. However, ALS has a better view of moderately sloping roadside features or v

vi

flat terrain such as of the pavement surface compared with ground-based systems depending on their scanner orientation. Features such as ditches and road barriers can be better viewed by ALS, whereas other systems are more appropriate to model the cliff slopes. Over the last decades, the development of LiDAR technology focused on increased measurements rates improved positional accuracy and angular precision and improvements in instrument functionality and easy to operate. The continuous developments of LiDAR will likely result in better road geometry modeling. This is because, with more accurate and denser point clouds, more detailed road features can be extracted and modeled. However, information such as road material types and surface roughness may require additional ancillary data such as multispectral and hyperspectral images. Overall, the ground-based systems are preferable to airborne systems for road geometry modeling due to the reduction in occlusion. In densely vegetated areas, the airborne LiDAR system is more likely to fail to collect detailed road features such as road markings. However, road geometry modeling using MLS and TLS data requires more sophisticated and efficient algorithms than those for ALS data. Chapter 2 presents a review of extraction and geometric modeling of road networks using LiDAR data. A brief explanation with graphical illustrations was given for geometric road models including their design and asset elements. This chapter also provided a novel classification of geometric road modeling according to the sensor type, preset objectives, and the technique used for data processing. Lastly, it discussed the current challenges and future trends in extraction and geometric modeling of road networks using LiDAR point clouds. The current algorithms do not meet the increasing demands and requirements of industrial applications. One key solution to this challenge was the automation of extraction and geometric modeling of road networks. However, most of the modern roads are complex in geometry, which by using automatic data processing methods; the detection accuracy is often not satisfactory. In addition, the presence of occlusions due to cars, trees, and buildings generates extra challenges and degrades the accuracy of the models. As a result, the future directions of extraction and geometric modeling of road networks should focus on developing models that can best solve the above issues. Chapter 3 discusses the recent technologies in mobile mapping systems that have enabled the rapid and cost-effective data acquisition on road corridors. This chapter presents a novel hierarchical strategy for the semiautomatic extraction of geometric road parameters from mobile laser scanning data. The hierarchical strategy included mean shift segmentation, particle swarm optimization, support vector machine classification, and principal component analysis. Accuracy assessments showed that the lowest and highest errors of the slope parameter were 1. 15 and 28.57%, respectively, whereas for the superelevation parameter was 0.96 and −13.50 %, respectively. These values of errors and comparative studies demonstrate the effectiveness of the method and indicate that the proposed model can offer satisfactory results. Chapter 4 presents optimization of parameters in extracting features from laser scanning data. Optimization is one of the essential tasks in machine learning algorithms. Classification algorithms, such as SVM and ensemble tree, require several user-defined parameters for model calibration and data classification. In this study, Taguchi-based optimization was implemented to select the best combination of user-defined parameters for the SVM and ensemble tree algorithms to extract road networks from LiDAR and aerial orthophoto data. The accuracy matrices indicated that the AUC values for the SVM and ensemble tree methods were 0.71 and 0.89, respectively, when the default parameters were used. Meanwhile, when the optimized parameters were used in classification, the AUC values of the SVM and ensemble tree methods were 0.88 and 0.95, respectively. This research demonstrates that optimizing the user-defined parameters of classification methods is necessary to improve accuracy and quality in detecting and classifying road networks from LiDAR data. In this study, the selected algorithms were trained using randomly selected samples and then tested on raster images. Further evaluation of the optimization-based classification is required to measure the efficiency and reliability of the approach. In addition, testing the transferability of the developed models on different study areas and datasets is a dynamic research trend.

Preface

Preface

vii

Chapter 5 presents an accurate approach for extraction of highway information from remote sensing data. This is significant for various applications such as traffic accident modeling, navigation, intelligent transportation systems, and natural hazard assessments. One of the conventional technique used for automatic highway extraction is machine learning. Despite several machine learning algorithms have been tested and tried in the recent years; however, there is no common agreement has been made which method performs better and spatially transferable. Therefore, this paper contributes in evaluating several machine learning algorithms (i.e. support vector machine, logistic regression, neural network, and decision tree) for automatic highway extraction from high-resolution airborne LiDAR data. Based on the comparative study performed, the best among studied machine learning algorithms was identified and used in an integrated GIS workflow for automatic highway extraction. The advanced GIS integrated workflow is an efficient model that could be applied to most commercial and open-source GIS software. Among the studied machine learning algorithms, multilayer perceptron and decision tree algorithms showed the best overall accuracy tested on randomly selected sampling data. However, when the transferability of the models investigated, logistic regression found to be the optimal algorithm for highway extraction from LiDAR data. In addition, although support vector machine produced high overall accuracy (90.19%) on sampling data, the model produced low-quality classification when applied to raster data. Thus, it suffers from model transferability issues. The quantitative evaluation showed that the logistic regression model could extract highway features from LiDAR data with 85.43% for completeness measure, 76.70%, and 67.82% for correctness and quality measures, respectively. The result of this study provides a clear guideline for other researchers to develop more advanced and automatic GIS models for accurate extraction of highways from LiDAR data. Chapter 6 discusses the several factors that contribute to road traffic accidents including human, vehicle, road geometry, and environmental factors. This chapter discusses the impact of roadside features (e.g., trees, access points, median, and shoulder) that were extracted from a mobile laser scanning (MLS) data on the injury severity of road traffic accidents. The analysis and the discussions were based on a case study of North–South Expressway (NSE) in Malaysia. The proposed methodology for roadside feature delineation was based on an object-based analysis comprised of multiresolution segmentation, optimization of features by random forest, and classification with support vector machines. Logistic regression (LR) model was utilized to analyze the correlation between the extracted features and the accident records from 2009 to 2015. The results of this study show that the proposed methodology could extract the roadside features from the MLS data with an overall accuracy of 86.45%. In addition, the estimated errors in calculating the road median, road width, and road shoulder width were 0.15, 0.31, and 0.45 m, respectively. On the other hand, results of LR revealed that shoulder width, high density of trees, poor lighting conditions, and involving motorcycle in crashes increase the injury severity of the accidents. Therefore, to improve the road safety in the focused area, this study suggests that these factors should be considered in road maintenance, safety management, planning, and other transportation projects. Chapter 7 presents a novel GIS-based model for automatic identification of road geometry in vector data. Some geospatial applications such as road safety assessment, car navigation, and updating digital road maps usually require road geometry information. In this research, a new model based on geographic information system (GIS) was proposed to automate the process of identifying road geometry in vector polylines. The proposed model first applies a Bezier interpolation to smooth the polylines for better cartographic representation. Then, the polylines were converted into raster data at 0.5 m spatial resolution. This data conversion enabled to convert the polylines to a set of points. After that, three geometry predictors were estimated from the set of points, point density, length of a line segment, and a cumulative angle between five consecutive points. Finally, the geometry predictors were used to predict the road geometry using three classification methods, namely support vector machine (SVM), decision tree (DT), and logistic regression (LR). The results show that the proposed model

viii

could identify road geometry in vector data. Overall accuracies of the three classification methods were 87.4, 89.7, and 87.2%, respectively; DT is the best. Chapter 8 reviews methods used to model the frequency and injury severity of traffic accidents. First, a general overview of modeling of traffic accidents was given. Second, road accident setting was described, which included configurations of road accidents across different studies. Some works researched traffic accidents on traffic signals, plaza tolls, whereas others focused on highways and main roads. Third, model factors for predicting the frequency and injury severity of traffic accidents were reviewed and discussed. Briefly, the selection of model factors depends on the purpose of research, the data availability, and the definition of the accident frequency or injury severity. In general, the reviewed studies showed that the model factors used for predicting road accident severity are usually more in number as compared to model factors used for predicting road accident frequency. After that, since most of the works showed that traffic accident data could be modeled by two primary models, statistical and computational intelligence, they were reviewed and compared considering their differences, similarities, advantages, and disadvantages. Finally, NN and DL models were further examined in details reviewing their concepts, use, the purpose of usage, and types. Besides the traditional NN model, RNN and CNN models were also discussed. In general, the accuracy of NN model decreases in particular with the increase in complexity of prediction, as is the case with road accident severity. The predictive accuracy of NN for accidents severity can be increased with the help of fusion algorithms and a clustering method. In addition, using a larger number of input nodes in the NN structure and more hidden layer, the predictive accuracy of NN is reported to be increased. Chapter 9 presents a comparative assessment of neural networks and support vector machines in modeling traffic accident severity using actual reported causes. With the significant increase in urbanization and number of registered vehicles worldwide, traffic accidents are a global concern. This chapter presents an evaluation of deep neural networks (DNNs) and support vector machines (SVMs) for traffic accident severity modeling using reported causes. A case study of Malaysian North–South Expressway (NSE) was presented. Accident data for the period of 2009–2015 consisting of 1138 observations were acquired for the study area. The data contained seven explanatory variables (vehicle type, accident cause, collision type, lighting condition, zone, bound, and road surface condition). The data were split into three subsets as follow: (1) training set, which contained 70% of the entire data; (2) validation set contained 15% of the data; and (3) testing set contained the remaining 15%. Using the training and validation datasets, the DNN and SVM models with optimum hyperparameters were built. After that, the trained models were tested using the testing dataset. The results indicated that the linear SVM outperformed other SVM models and NN models. The highest accuracy on testing dataset was 71.34% obtained by the linear SVM model. In addition, the best DNN model obtained 70.67%. Furthermore, the random forest analysis showed that the reported accident cause and vehicle type are the two most influential factors increasing the severity of accidents in the study area. Modeling the injury severity of road traffic accidents is a challenging yet essential task in transportation management. Predictive models should be powerful enough to make accurate forecasts and generalize well on different data subsets for a given area. Chapter 10 presents a new hybrid algorithm based on extreme gradient boosting (XGBoost) and deep neural networks (DNNs) for predicting the injury severity of road traffic accidents. The proposed model workflow includes handling imbalance data, optimizing the input factors, fine-tuning the model’s hyperparameters by a grid search method, and combining two dominant models XGBoost and DNN in a single hybrid model to improve the generalization performance. The findings show that the handling imbalance data with the SMOTEENN method can improve the model performance significantly (*by 12%). In addition, feature selection and optimization of the hyperparameters could help improving the model regarding the prediction accuracy and generalization. The new hybrid model had an average accuracy of 0.96, which outperformed many other machine learning models like K-nearest neighbors, decision trees, random forest,

Preface

Preface

ix

multilayer perceptron, support vector machine, naïve Bayes, Gaussian process classifier, and quadratic discriminant analysis methods. On the other hand, the XGBoost model with optimum features also performed well (0.95) slightly worse than the hybrid model. The new models allow better road safety assessment, and with their high predictive and generalization capacity, they provide a better understanding of road accidents on highways. Information created from such advanced models can help decision makers to take better decisions eliminating the significant losses due to accidents in future. Traffic accidents are becoming a growing concern due to the increasing number of casualties and losses in economic activities. Driver, highway, vehicle, accident characteristics, and climatic factors are some of the factors that determine the severity level. These factors are often analyzed using statistical and machine learning models. In recent years, new advances in computer hardware/software engineering and large datasets have assisted deep learning to win numerous contests in pattern recognition, machine learning, and remote sensing. Deep learning enables hierarchal learning by computational models of data using multiple abstraction levels at different processing layers. Chapter 11 investigates the power of deep learning in predicting the severity of injuries when accidents occur due to traffic on Malaysian highways. Three network architectures based on a simple feedforward neural networks (NNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs) were proposed and optimized through a grid search optimization to fine-tune the hyperparameters of the models that can best predict the outputs with less computational costs. The results showed that among the tested algorithms, the RNN model with an average accuracy of 73.76% outperformed the NN model (68.79%) and the CNN (70.30%) model based on a tenfold cross-validation approach. On the other hand, the sensitivity analysis indicated that the best optimization algorithm is ‘Nadam’ in all the three network architectures. In addition, the best batch size for the NN and RNN was determined to be 4 and 8 for the CNN. The dropout with keep probability of 0.2 and 0.5 was found critical for the CNN and RNN models, respectively. This research has shown that deep learning models such as CNN and RNN provide additional information inherent in the raw data such as temporal and spatial correlations that outperform the traditional NN models regarding both accuracy and stability. Chapter 12 presents a novel deep learning model based on recurrent neural network (RNN) and transfer learning (TL) for predicting injury severity of motorcycle accidents on Malaysians expressways. Due to the limited available data (155 records) on motorcycle accidents, we first developed a model based on 970 accident records that accurately predicts vehicle accidents on Malaysians expressways. Then, this model was retrained on a small dataset of the motorcycle accidents by freezing the first few layers. In addition, the network architecture and its hyperparameters were optimized via grid search method. The best network comprised of one long short-term memory (LSTM) layer (128 hidden units) and two dense layers of 100 hidden units. Moreover, it also contained dropout layers with a probability of 0.2 to reduce the model complexity and improve the generalization performance of the model. The network was trained using RMSprop optimization method with learning rate and weight decay of 0.001 and 0.0001, respectively. The model was implemented using Google’s TensorFlow. The proposed model achieved a training and testing accuracy of 78.29 and 77.90%, respectively, and the TL could improve the accuracy by almost 10%. A comparative study showed that the RNN with TL method outperformed other machine learning models such as support vector machine, logistic regression, and random forest. The model also could explain the effects of accident-related factors on the injury severity outcomes by calculating factor’s coefficients by a connection weight method. The results suggested that collision type, accident cause, and lighting condition are the most influential factors for increasing the injury severity of motorcycle accidents on Malaysians expressway. Moreover, the proposed model can efficiently use the temporal and spatial structure of the accident data as additional information to improve the prediction performance, which other methods lack.

x

Preface

The publication of this book would not have been possible without an excellent cooperation from my colleagues at Springer Verlag, Germany. Special thanks to Dr. Nabil Khelifi for motivating and encouraging the first author to write this book. At Springer Verlag, the efforts from Reyhaneh Majidi is appreciable. Sydney, Australia April 2019

Biswajeet Pradhan

Contents

Part I 1

2

Road Geometry Modelling

Laser Scanning Technologies in Road Geometry Modeling . . 1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Laser Scanning Systems . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Airborne Laser Scanning . . . . . . . . . . . . . . . . . . 1.2.2 Mobile Laser Scanning . . . . . . . . . . . . . . . . . . . 1.2.3 Terrestrial Laser Scanning . . . . . . . . . . . . . . . . . 1.3 Description of Scanner Field of View . . . . . . . . . . . . . . . 1.4 Comparison of ALS, MLS, and TLS for Road Extraction and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

3 3 3 3 6 7 8

............ ............ ............

8 11 11

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

Road Geometric Modeling Using Laser Scanning Data: A Critical Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Road Geometric Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Delineation of Road Geometric Information: A Novel Classification for LiDAR Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Road Extraction Based on Types of Sensors . . . . . . . . . . . . . . . . . . . 2.4.1 Satellite Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Aerial Photographs and Unmanned Aerial Vehicles (UAV) . . 2.4.3 LiDAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5 Classification of Road Extraction According to the Preset Objective . . 2.5.1 Road Centerline Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.2 Road Extraction in Two Dimensions . . . . . . . . . . . . . . . . . . . 2.5.3 Road Reconstruction and Geometric Modeling . . . . . . . . . . . 2.6 Classification of Road Extraction According to the Technique Applied 2.6.1 Unsupervised/Supervised Classification . . . . . . . . . . . . . . . . . 2.6.2 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.3 Object-Based Image Analysis (OBIA) . . . . . . . . . . . . . . . . . . 2.6.4 Morphology and Filtrate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.5 Active Contour Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.6 Information Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6.7 Hierarchal and Hybrid Methods . . . . . . . . . . . . . . . . . . . . . . 2.7 Discussion and Current Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . 2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

... ... ...

15 15 15

. . . . . . . . . . . . . . . . . . . .

17 17 17 17 18 18 18 19 20 20 20 21 22 23 23 24 24 25 28 28

. . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . .

xi

xii

3

4

5

Contents

Road 3.1 3.2 3.3

Geometric Modeling Using a Novel Hierarchical Approach Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Previous Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 MS Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Optimization of Segmentation Parameters . . . . . . . . . 3.3.3 Classification of Segments Using SVM . . . . . . . . . . . 3.3.4 Extraction of Geometric Road Parameters Using PCA Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Accuracy Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Evaluation of Model Transferability . . . . . . . . . . . . . . . . . . . 3.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

33 33 33 34 35 36 37

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

37 37 42 42 44 44 45

. . . . . . . . . .

. . . . . . . . . .

47 47 48 48 49 49 50 50 50 51

.. ..

51 52

Optimizing Support Vector Machine and Ensemble Trees Using Taguchi Method for Road Extraction from LiDAR Data . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1 SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.2 Ensemble Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.3 Taguchi-Based Optimization of User-Defined Parameters . . . . . 4.2 Study Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Overall Workflow of the Study . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Data Preprocessing and Preparation . . . . . . . . . . . . . . . . . . . . 4.3.3 Optimization of User-Defined Parameters (Taguchi Method) . . 4.3.4 Training the SVM and Ensemble Tree Algorithms with Default and Optimized Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Results of the Classification Using the Default User-Defined Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Results of the Taguchi-Based Optimization . . . . . . . . . . . . . . . 4.4.3 Results of the Classification Using the Optimized User-Defined Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.4 Accuracy and Performance Assessments . . . . . . . . . . . . . . . . . 4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.. ..

52 53

. . . .

. . . .

56 56 58 58

An Integrated Machine Learning Approach for Automatic Highway Extraction from Airborne LiDAR Data and Orthophotos . . . . . . . . . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Previous Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Machine Learning Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1 Multilayer Perceptron Neural Networks (MLP) . . . . . . . . . . . 5.3.2 Support Vector Machine (SVM) . . . . . . . . . . . . . . . . . . . . . . 5.3.3 Logistic Regression (LR) . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.4 Decision Tree (DT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Study Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Data and Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.1 Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.2 Preparation of Input Attributes and Training/Testing Samples . 5.5.3 Proposed GIS Workflow for Automatic Highway Extraction .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

61 61 62 63 63 63 64 64 65 65 65 66 66

. . . . . . . . . . . . .

Contents

xiii

5.6

and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Proposed Models for Highway Extraction . . . . . . . . . . . Accuracy Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . Multilayer Perceptron . . . . . . . . . . . . . . . . . . . . . . . . . . Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . Applications on Raster Data and Models Transferability Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.6 Quantitative Evaluation of Road Extraction . . . . . . . . . . 5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

67 68 68 68 70

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

70 71 73 74

6

Effect of Roadside Features on Injury Severity of Traffic Accidents 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2.1 Overall Flowchart . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2.2 Data Preparation and Preprocessing . . . . . . . . . . . . . . . 6.2.3 Roadside Feature Extraction . . . . . . . . . . . . . . . . . . . . . 6.2.4 Feature Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2.5 Impact Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.2 Roadside Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.3 Results of LR Modeling . . . . . . . . . . . . . . . . . . . . . . . . 6.3.4 Results of Impact Assessment . . . . . . . . . . . . . . . . . . . . 6.3.5 Results of Feature Contributions . . . . . . . . . . . . . . . . . . 6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

77 77 78 78 78 78 81 81 81 81 81 83 83 84 84 85

7

Novel GIS-Based Model for Automatic Identification of Road Geometry in Vector Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Airborne Laser Scanning . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.2 Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.3 Extraction of Road Geometric Design Parameters . . . . . . . . . 7.3 Results and Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 Application of Bezier Interpolation on Road Polylines . . . . . . 7.3.2 Results of Road Geometry Identified in Vector Data . . . . . . . 7.3.3 Characteristics of Identified Curves . . . . . . . . . . . . . . . . . . . . 7.3.4 Validation of Different Classification Methods . . . . . . . . . . . . 7.3.5 Results of Calculating NSE Design Parameters . . . . . . . . . . . 7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

87 87 87 87 88 88 89 89 90 91 91 91 94 94

Networks . . . . . . . . . . . .

97 97 97 98 98 99 100

Part II 8

Results 5.6.1 5.6.2 5.6.3 5.6.4 5.6.5

Modeling Road Traffic Accidents

Review of Traffic Accident Predictions with Neural 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Background of Neural Networks . . . . . . . . . . . 8.3 The Common Types of Neural Networks . . . . . 8.3.1 Feedforward Neural Networks (FNNs) 8.3.2 Convolutional Neural Network (CNN) 8.3.3 Recurrent Neural Network (RNN) . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

xiv

Contents

8.4 8.5

Road Accident Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Modeling Road Traffic Accidents . . . . . . . . . . . . . . . . . . . . . . . . 8.5.1 Model Factors for Predicting Road Accident Frequency . . 8.5.2 Model Factors for Predicting Road Accident Severity . . . 8.5.3 Frequency Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6 Injury Severity Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.7 The Use of Neural Networks for Road Traffic Accident Modeling 8.8 The Rationale of Using NN . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.9 Comparison of NN and Statistical Models for Traffic Accident Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.10 Predictive Performance of Different Types of NN . . . . . . . . . . . . 8.11 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

101 101 101 102 102 103 104 104

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

104 107 108 108

Modeling Traffic Accident Severity Using Neural Networks and Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3 Study Area and Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.1 DNNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3.2 SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.1 Results of DNN Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.2 Results of SVM Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.3 Results of RF Factor Importance . . . . . . . . . . . . . . . . . . . . . 9.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

111 111 111 112 112 112 113 113 114 114 116 116

10 Predicting Injury Severity of Road Traffic Accidents Using a Hybrid Extreme Gradient Boosting and Deep Neural Network Approach . . . 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Previous Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.1 Data and Application Site . . . . . . . . . . . . . . . . . . . . . . . 10.3.2 Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.3.3 Evaluation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4.1 Results of Handling Imbalanced Data . . . . . . . . . . . . . . . 10.4.2 Results of Feature Importance . . . . . . . . . . . . . . . . . . . . 10.4.3 Performance of Hybrid Model . . . . . . . . . . . . . . . . . . . . 10.4.4 Comparison with Other Methods . . . . . . . . . . . . . . . . . . 10.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

119 119 119 120 120 120 122 123 123 124 124 124 126 127

11 Applications of Deep Learning in Severity Accidents . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Introduction . . . . . . . . . . . . . . . . . . . 11.2 Deep Learning Models . . . . . . . . . . . 11.2.1 Feedforward NNs . . . . . . . . 11.2.2 CNNs . . . . . . . . . . . . . . . . . 11.2.3 RNNs . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

129 129 129 129 130 131

Prediction of Traffic . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

Contents

xv

11.3 Proposed Models . . . . . . . . . . . . . . . . . . . . . . 11.4 Experimental Results . . . . . . . . . . . . . . . . . . . 11.4.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . 11.4.2 Model Performance . . . . . . . . . . . . . . 11.4.3 Optimization and Sensitivity Analysis . 11.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

131 132 133 133 133 137 138 138

12 Forecasting Severity of Motorcycle Crashes Using Transfer Learning . 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 RNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.3 TL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4 Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4.1 Network Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4.2 Network Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4.3 Data Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4.4 Hyperparameter Optimization . . . . . . . . . . . . . . . . . . . . . . 12.4.5 Mitigating Overfitting Problem . . . . . . . . . . . . . . . . . . . . . 12.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5.2 Training from Scratch Versus TL . . . . . . . . . . . . . . . . . . . 12.5.3 Effects of Batch Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.5.4 Effects of Data Transformation . . . . . . . . . . . . . . . . . . . . . 12.5.5 Effects of Network Hyperparameters . . . . . . . . . . . . . . . . . 12.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.6.1 Time Complexity of the Model . . . . . . . . . . . . . . . . . . . . 12.6.2 Model Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.6.3 Importance of Accident-Related Factors . . . . . . . . . . . . . . 12.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . .

141 141 142 142 143 144 144 145 145 146 146 146 149 151 152 152 153 153 154 154 155 156

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

Part I Road Geometry Modelling

1

Laser Scanning Technologies in Road Geometry Modeling

1.1

Background

Recent developments in laser scanning technology have improved 3D spatial data acquisition in road environments. These technological developments have provided rapid and cost-effective data acquisition for road corridors and surrounding areas (Guan et al. 2014; Li et al. 2016; Li and He 2016; Lin et al. 2011). Several methods have been proposed for the delineation of geometric road information from laser scanning data. Road geometric information includes road width, cross section, and superelevation and involves the number of road lanes and vertical and horizontal curves. Road geometric parameters are essential for various applications, such as city and road environment modeling (Goulette et al. 2006; Kukko et al. 2009), evaluating traffic noise pollution (Fiedler and Zannin 2015), road safety (Camacho-Torregrosa et al. 2013), and optimizing transportation facilities (Kang et al. 2013). Dense point clouds acquired by laser scanning systems [or light detection and ranging (LiDAR)] facilitate accurate and detailed extraction of geometric road parameters through several methodological workflows. Traditional road surveying and mapping activities utilize aerial photographs with an image interpreter by manually digitizing roads. This method is time-consuming and provides poor quality because the process is often affected by human errors, especially in large projects (more than 5 km2). Field surveying is another widely applied practice for updating road information. The “total station” or “digital levelling” systems are mainly used in this method. Although accurate, this approach is costly and time-consuming for large-scale projects. Another method is LiDAR, a standard remote-sensing technique used for updating road databases because of its efficiency and high accuracy over large areas (White et al. 2010). Pereira and Janssen (1999) confirmed that a digital surface model (DSM) derived from laser measurements with an average point density of 4 points/m2 exhibited sufficient quality to represent terrain relief for preliminary road planning and design. This chapter presents an overview of laser scanning technology for road modeling. First, laser scanning systems,

namely airborne, mobile, and terrestrial LiDAR systems, are introduced. The basic principles, advantages, and disadvantages of these technologies are explained with graphical illustrations. Finally, this chapter provides a comparison of airborne, mobile, and terrestrial LiDAR systems regarding their efficiency and performance in road environment modeling.

1.2

Laser Scanning Systems

Laser scanning systems are based on laser ranging, which measures the distance between a sensor and a target using the elapsed time between the emission of an infrared pulse (frequency of 150 kHz or higher) and the detection of a reflected return (Baltsavias 1999; White et al. 2010). Extant literature classified laser scanning systems as either discrete return or full-waveform recording (Wulder et al. 2012). Discrete laser scanning systems record single or multiple returns from a given laser pulse. By contrast, full-waveform laser scanning systems record the entire reflected energy from a return, thereby resulting in complete sub-meter vertical profiles of a target. From the perspective of geometric road modeling, discrete return laser scanning systems are usually preferred because laser pulses cannot penetrate road surfaces to record the complete vertical profile of the target. Another type of classification for laser scanning systems is based on the type of platform. Based on the platform type, the following three types of laser scanning systems are classified: an airborne laser scanning system, a mobile laser scanning system, and a terrestrial laser scanning system. The following sections describe each of these systems in detail.

1.2.1 Airborne Laser Scanning Airborne laser scanning (ALS) is terrain and urban scanning technology applied to acquire highly accurate bathymetric and topographic data. Such data are frequently collected

© Springer Nature Switzerland AG 2020 B. Pradhan and M. Ibrahim Sameen, Laser Scanning Systems in Highway and Safety Assessment, Advances in Science, Technology & Innovation, https://doi.org/10.1007/978-3-030-10374-3_1

3

4

using a fixed-wing aircraft or helicopter. Dense point clouds can be gathered using helicopter-based LiDAR because of the low aircraft speed. A typical ALS system consists of three complementary technologies (Fig. 1.1): a laser to measure the distance to the target; an inertial measurement unit (IMU) to record the pitch, roll, and yaw of the platform; and a kinematic global positioning system (e.g., GPS and GLONASS—global navigation satellite system). The absolute position of a reflecting surface can be ascertained by combining information from each of these technologies using accurate time referencing (Wulder et al. 2012). The result is a record of dense, 3D point clouds that represent the ground surface topography of the surveyed terrain. The point clouds describe the 3D topographic profile of the Earth’s surface. The advantages of ALS include lack of effects of relief displacement, penetration of tree canopy, and insensitivity to lighting conditions (Yan et al. 2015). Applications of ALS in transportation involve road location and design, road extraction and geometry modeling (Hui et al. 2016; Ferraz et al. 2016; Sameen and Pradhan 2017a, b, c), road safety (Jalayer et al. 2015), traffic flow prediction (Liu et al. 2016; Zhang et al. 2014), and road inventory mapping (He et al. 2017).

1.2.1.1 Height Features of ALS Data LiDAR height information is regarded as a useful feature for delineation of elevated objects from bare earth features (Idrees and Pradhan 2016, 2018; Idrees et al 2016; Fanos et al. 2016, 2018; Fanos and Pradhan 2018; Abdulwahid and Pradhan 2017; Pradhan et al. 2016; Sameen et al. 2017; Sameen and Pradhan 2017a, b). Users interpolate the 3D point clouds (Fig. 1.2a) to produce a digital surface model (DSM) and digital elevation model (DEM) that can increase the signature separability among different land cover classes. By incorporating ALS-derived height features on Fig. 1.1 Platform of an airborne laser scanning system

1

Laser Scanning Technologies in Road Geometry Modeling

multispectral images, accuracy improvements of 5–6% were reported in several studies (Hartfield et al. 2011). Charaniya et al. (2004) acquired a 10 m DEM provided by the United States Geological Survey (USGS) and an ALS-derived DSM to produce a normalized height (nDSM). With the mature development of point cloud filtering techniques, the bare ground layer can be generated from the existing point clouds along with the nDSM. Such nDSM has demonstrated its practicality in improving classification accuracy (Brennan and Webster 2006; Hartfield et al. 2011). Data resolution refers to the resolution of interpolated surfaces, which is always found while using discrete return LiDAR data. Note that most of the studies used 0.5–1 m data resolution, which seems to be a good tradeoff between the mean point spacing and point density (Yan et al. 2015).

1.2.1.2 Laser Intensity Radiometric component of LiDAR data or laser intensity (Fig. 1.2b) can serve as an additional feature in the classification domain (Yan et al. 2015). Laser intensity represents the peak amplitudes recorded in the laser beam return backscattered from the illuminated object, where the intensity is usually linearized into an 8–12-bit data scale. Laser intensity depends on the range, incidence angle of the laser beam, and surface characteristics (Kumar et al. 2014). In addition, Charaniya et al. (2004) concluded that laser intensity demonstrated a critical role in the road and low vegetation when using only height information presents difficulties. Intensity data can also be integrated with aerial orthophotos to improve the classification of land cover and road features. Although the functionality of intensity data has proven to be efficient, it has drawbacks as well. One of its significant disadvantages is that the data are noisy. For example, a significant variance in the tree features occurs because of the irregular geometry of the canopy surface

1.2 Laser Scanning Systems

5

Fig. 1.2 Airborne LiDAR data. a LiDAR point clouds colored by elevation attributes. Low elevations are shown in blue, whereas high elevations are in red. b LiDAR intensity data. Low intensities are in

dark colors, and high intensities are in bright colors. c True color aerial orthophoto acquired by the same airborne LiDAR system in the same data acquisition mission

(Yan et al. 2012; Minh and Hien 2011). The limitations of intensity data degrade the accuracy and performance of classification. Another disadvantage of such data is radiometric misalignment, which is found in overlapping LiDAR data strips (Brennan and Webster 2006). Therefore, preprocessing and radiometric calibration for intensity data before classification and feature extraction are crucial.

1.2.1.4 Specifications of ALS Sensors in Road Geometric Modeling Different LiDAR specifications were used in data acquisition on-road environments. Table 1.1 summarizes the common sensor types reported in the road extraction literature. The average point density starts from 1.6 to 4 points/m2. The typical flight height was found to be *1200 m. Some of the studies employed different flight heights due to the type of topography and sensor resolution. The common sensor types used for road data collection are Leica and Optech. The Leica ALS50-II is a compact, ALS system designed to acquire information topographically and return signal intensity data from various airborne platforms. The system computes the data using laser range and returns signal intensity measurements recorded in-flight, along with the position and altitude data taken from airborne GPS and IMU systems. Conversely, Optech ALTM 3100EA sets the global standard for airborne laser terrain mapping with data

1.2.1.3 Aerial Orthophotos Digital cameras are mounted on laser scanning systems to capture very high-resolution images of point clouds. Images collected by laser scanning systems represent valuable data, which can be integrated with point clouds to extract useful information and recognize various objects, including different types of roads. Figure 1.2c shows an example of an aerial orthophoto collected by ALS over the University Putra Malaysia (UPM); the orthophoto displays a mixture of land covers, such as roads, buildings, vegetation, and water bodies.

Table 1.1 Performance specifications of common airborne laser scanning systems used for road extraction and geometric road modeling in previous studies

Sensor type

Reported point density

Flight altitude

Leica ALS50

4 points/m2 (average)

1200–1500 m

Optech 3100EA

2–4 points/m2 (average)

1200 m

Leica ALS50-II

0.5 points/m2 (minimum)

1800 m

Optech ALTM3070

1.6 points/m2 (average)

Not reported, but the operational altitude of the device is 80–3500 m

6

acquisition rates of up to 100,000 pulses/s and innovative features, such as intensity, full-waveform digitization, simultaneous first/last pulse measurement, roll compensation, and digital camera integration.

1.2.2 Mobile Laser Scanning The first mobile mapping system, GPS-Van, was developed in the early 1990s by the Center for Mapping at Ohio State University (Puente et al. 2013). A typical mobile-mapping system consists of three components: mapping sensors, a positioning and navigation unit for spatial referencing, and a time referencing unit (Fig. 1.3). Mobile laser scanning (MLS) involves mobile mapping systems that use LiDAR technology as the central unit. MLS shares many features with ALS, one of which involves the fundamentals of laser distance measurement and scanning. Moreover, data processing workflows are very similar (or nearly identical) in both cases. By contrast, these systems differ about typical project size and obtainable accuracy and resolution, among other factors. ALS systems almost exclusively use the pulse time-of-flight measurement principle for ranging, and they have been widely utilized for the generation of bare-earth DEM. Compared with ALS, locating the scanner on a mobile ground platform provides distinct advantages for the capture of discrete objects from multiple angles. Any MLS system integrates several sub-systems, including digital frame cameras, a laser scanner, an IMU in combination with a GNSS, and a control unit that operates all of these components, synchronizes measurement acquisition, and records the collected data (Fig. 1.3). Recent MLS systems employ Fig. 1.3 Platform of a mobile laser scanning system (e.g., Topcon IP-S3 HD 3D mobile mapping system)

1

Laser Scanning Technologies in Road Geometry Modeling

two techniques for range measurements: time-of-flight (TOF) and phase shift. The TOF scanner sends a short laser pulse to the target, and the time difference between the emitted and received pulses is used to determine the range. In comparison, phase-based laser scanners employed the phase difference between the emitted and received backscattered signal of an amplitude-modulated continuous wave to determine the range. Phase-shift laser scanners are accurate, but their measurement range is shorter than that of TOF. Mobile mapping systems were invented when GPS became available. Its basic concept is to utilize sensors installed on a moving vehicle, such as a car, truck, and rail car, to acquire geospatial data in a highly efficient way over transportation corridors at normal travel speeds. For example, Fig. 1.4 shows colored point clouds over a highway section captured by a Riegl VZ-2000 mobile scanner and a Nikon D800 camera (30 megapixels) on top of a moving vehicle at a speed of 30–40 km/h during data acquisition. The details allowed extraction and characterization of road geometry and roadside features, including a road sign, trees, and road barriers. The two most critical components of mobile mapping systems are direct platform georeferencing and digital imaging. In addition, digital workflow and a high level of automation were presumed to be present. Initially, achieving all of the objectives was difficult because neither digital imaging sensors nor processing software was available. After the slow start, however, technological advancements quickly propelled mobile mapping forward. To reflect the significant performance improvements, the term “mobile mapping technology” was introduced about a decade ago after the Fourth International Symposium on Mobile Mapping

1.2 Laser Scanning Systems

7

Fig. 1.4 An example of MLS point clouds acquired by a Riegl VZ-2000 mobile scanner and a Nikon D800 camera (30 megapixels) for a highway section

Technology (Tao and Li 2007). MLS applications in transportation include road location and design, road extraction and geometry modeling (Yang et al. 2013; Kumar et al. 2013), highway safety, and road inventory mapping (Pu et al. 2011; Yu et al. 2015; Gong et al. 2012; Guan et al. 2015).

1.2.3 Terrestrial Laser Scanning Like ALS and MLS, static terrestrial laser scanning (TLS) provides accurate 3D point clouds that can be used to extract geometric road features and road furniture. TLS systems are often characterized by their scanning patterns (camera, hybrid, and panoramic) for data acquisition or their fundamental range measurement principle (phase- or pulse-based). This section describes the working principle of these systems. TLS is used for transportation applications, such as traffic safety analysis (Pagounis et al. 2007), object classification in railway environments (Arastounia 2012), and highway tunnels (Gikas 2012). Figure 1.5 illustrates the schematic operation principle of a typical TLS system. The working principle of TLS depends on recurrent measurements of the slope range (HD) measured by an electronic distance measurement (EDM) device at known horizontal (h) and vertical (£) angular intervals. This process outputs dense points with spherical polar coordinates in the field of view of the scanner. Although a robotic total station can measure very specific features, TLS systems with angular measurements over a predefined angular range can measure denser points in a shorter period. Furthermore, TLS systems can measure the

reflection intensity of the targets in sight. The main benefit of reflection intensity is the strength of the returned laser beam from the different surface materials. Although reflection intensity is affected by the incidence angle and the distance between the scanner and the target, it provides several advantages, such as the interpretation of physical characteristics (surface roughness and material type). Moreover, TLS systems are commonly equipped with digital cameras used for mesh-texturing imagery. The laser scanners identify the range between the sensor and the targets by measuring the amount of time required for the laser beam to travel to the target and return, as represented by the following expression: r¼

c Dt  ; n 2

ð1:1Þ

where c is the speed of light (3  108 m=s), n is the reflective index of the light traveling medium, and Dt is the time difference between the deployment and return of the laser beam. As mentioned, TLS systems can be classified according to their fundamental range measurement principle, Dt, which can be ascertained by either pulse-based or phase-based methods. In the former, the laser system emits a short pulse (1–10 ns) from the laser source and measures the trip travel time (Dt) by an accurate internal timing circuit. With these systems, the range can be quantified with a precision of approximately 3–15 mm (Vosselman and Maas 2010). By contrast, phase-based TLS systems use a continuous wave modulation where a continuous stream of laser light is modulated in amplitude. Dt is then computed by measuring

8

1

Laser Scanning Technologies in Road Geometry Modeling

Fig. 1.5 Schematic operation principle of a TLS system

the phase difference between the emitted and reflected laser beams. The phase difference is calculated as follows: D£ km Dt ¼  ; 2p c

ð1:2Þ

where D£ is the phase difference, and km is the amplitude modulation wavelength. The range of phase-based systems is determined by substituting Eq. 1.2 into Eq. 1.1. These systems often have better range precision of approximately 1–10 mm. Regarding the maximum range, pulse-based systems are capable of measuring a range of several kilometers, whereas phase-based systems are limited to several hundred meters depending on the power requirements for generating a continuous wave and avoiding phase ambiguity (Vosselman and Maas 2010).

1.3

Description of Scanner Field of View

As discussed, ALS measures the distance or range from the sensor to the ground objects by firing laser pulses that reach the ground by traveling through the atmosphere at the velocity of light. The fired pulses are often sent successfully in the field of view (FOV) (Fig. 1.6a). The FOV is equally divided on both sides on the nadir direction at the laser source. The FOV is also designated as the scan angle and is expressed by 2£. This angle, in addition to the flight height (H) with reference to a datum, provides the “swath” (Bs ) or the width of a flight strip on the ground. That width can be calculated using Eq. (1.3) as follows:

Bs ¼ 2H  tan £:

ð1:3Þ

However, in MLS, the FOV commands the extent of the area that can be covered in a single pass of the collection vehicle. The FOV depends on the number of sensors used. When two sensors oriented in an “X” pattern are utilized, a scanning FOV of 360° can be achieved (Fig. 1.6b). The first sensor captures the front face of an object during the approach, and the second monitors the back face during the departure. In TLS, the laser source is fixed on a tripod, and the laser beam is directed in two directions, allowing for the generation of 3D point clouds for the surface that must be scanned. TLS scanners have three scan patterns: camera (Fig. 1.7a), hybrid (Fig. 1.7b), and panoramic (Fig. 1.7c). In the camera pattern, the laser beam is directed in two directions through two mirrors that are orthogonally mounted and independently rotated. For the hybrid pattern, a single mirror is rotated in one direction along with rotating the entire scanner through the FOV, providing a full 360° scan around the vertical axis. Unlike hybrid scanners, panoramic scanners collect an entire “dome” of measurements covering the upward direction.

1.4

Comparison of ALS, MLS, and TLS for Road Extraction and Modeling

Laser scanning technology offers many benefits to transportation agencies, particularly for geometric road modeling. The corresponding scanning systems share a few

1.4 Comparison of ALS, MLS, and TLS for Road Extraction and Modeling

9

Fig. 1.6 Illustrations for FOV of a ALS and b MLS systems

Fig. 1.7 Visual illustration of three TLS scanning patterns: a camera, b hybrid, and c panoramic

similarities, such as hardware components (GNSS, IMU, and laser scanner), collecting point clouds, providing laser return intensity, and supporting cameras to collect color information on the surveyed surfaces. However, they differ in many aspects as explained in the following points (see Table 1.2 for a summary). Regarding safety, MLS has increased safety compared with ALS, static TLS, and traditional surveying methods because all its measurements are taken from inside a moving vehicle (Yen et al. 2011). This property precludes driver distraction due to survey instruments, provides flexibility to deal with precarious situations, and facilitates movement with the traffic flow, thereby eliminating the need to close the roadway. Glennie (2009) provided an efficiency comparison between MLS and static TLS over a four-mile section of the interstate road. The total scanning time with Terrapoint’s mobile mapping system was 1.5 h while the roadway

remained fully open during the survey, which is significantly less than the time needed for static TLS or the “total station” system (*10 working days) (Zampa and Conforti 2009). Regarding the operation process, the usage of TLS systems is cheaper and less complex because ALS requires integrated IMU and GNSS sensors for accurate point georeferencing, whereas MLS necessitates the estimation of the motion of the scanner atop the moving vehicle. Given the larger altitude of flight of ALS, its point density is more uniform than that of other systems, depending on terrain elevation variations. MLS and TLS systems collect points that are close to the scanner and less densely farther from the scanner path (Williams et al. 2013). For the view angle, depending on the scanner orientation, ALS has a better view of moderately sloping roadside features or flat terrains, such as that of the pavement surface, compared with MLS and static TLS. Features, including

10 Table 1.2 Performance comparison of ALS, MLS, and TLS for road extraction and modeling [modified after Williams et al. (2013)]

1

Laser Scanning Technologies in Road Geometry Modeling

ALS

MLS

TLS

Direct view of pavement and building tops

Good view of pavement, unable to capture building tops

Good view of pavement with details

Oblique view of vertical faces

Direct view of vertical faces

Direct view of vertical faces with more flexibility

Fast coverage

Slow coverage

Slower coverage

Large footprint

Small footprint

Small footprint

Far-range travel

Short-range travel

Short-to-moderate range travel

Not limited to the area visible from the roadway

Limited to objects close to and visible from the roadway

Limited to objects close to the roadway

Low point density (1–60 point/m2)

High point density (100 points/m2)

Very high point density (500 points/m2)

Limited options for setup locations

Good options for setup locations

Better options for setup locations

Difficult to operate and requires adequate training

Difficult to operate but easier than ALS as a pilot is not needed

Easy to operate with less training

Provides a low level of details

Provides a high level of details

Provides the highest level of details

Low accuracy and resolution of road features

High accuracy and resolution of road features

Higher accuracy and resolution of road features

Highest cost-effectiveness

High cost-effectiveness

Low cost-effectiveness

ditches and road barriers, can be better viewed by ALS, whereas other systems are more appropriate for modeling cliff slopes. The laser footprint on the ground is much larger for ALS than for other systems, thereby leading to more horizontal positioning uncertainty with ALS. All three systems have error sources. Such sources include the GNSS, IMU, and laser footprint (for ALS) and GNSS measurements (for MLS and TLS). MLS and static TLS can capture surfaces underneath bridges and in tunnels, a feature that ALS lacks. By contrast, ground-based laser scanning systems are insufficient in collecting data within a short and moderate range (100– 1000 m) of roadways and surrounding features. Compared with MLS and ALS, static TLS provides more options for setup locations, including positions away from the road. Users can also determine the desired resolution with a single configuration. This attribute enables static scanning to obtain a higher resolution on objects, such as targets. Higher accuracies and resolutions can be achieved because the platform is stationary (Williams et al. 2013). MLS and TLS offer significantly improved horizontal accuracy because of their looking angles (