Data-driven research on vehicles' risky lane-changing manoeuvre

• Main Author: Chen, Tianyi
• Other Authors: Wong Yiik Diew
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Civil engineering::Transportation
• Online Access: https://hdl.handle.net/10356/147066
• Institution: Nanyang Technological University

Risky Lane-changing (LC) manoeuvre can lead to serious damage to property and great loss of life. Hence, it is of great significance to investigate into risky LC manoeuvre and improve road traffic safety. However, the research on risky LC manoeuvre faces the challenges of rare crash data, behavioural complexity, multiple contributing factors, and sophisticated causal mechanism. This research is undertaken to resolve various challenges in four phases (i.e., Phases 1 to 4), which are briefly summarised as below. Next Generation Simulation (NGSIM) dataset is used to validate the studies in Phases 1 to 3, while in Phase 4 the Second Highway Research Program (SHRP2) Naturalistic Driving Study (NDS) dataset is applied for method validation. The study in Phase 1 develops a general framework for key feature selection and risk prediction of a car’s LC behaviour on the highway. This study also proposes a method to comprehensively evaluate resampling methods which alleviate class imbalance problem. As an efficient machine learning technique, Random Forest classifier is applied for feature selection and risk prediction, which achieves satisfactory prediction performance based on the NGSIM dataset as validation. The results of feature selection imply that the behaviours of the surrounding cars shall be paid more attention when investigating the safety of an LC event. The study in Phase 2 aims to develop a pre-emptive LC risk level prediction (P-LRLP) method to estimate the crash risk level of an LC event in advance before the LC car completes the LC manoeuvre. This study proposes a resampling method, namely, ENN-SMOTE-Tomek Link (EST), and an algorithm to evaluate feature selection method. The validation results manifest that the EST can achieve promising resampling performance, and Random Forest classifier as an embedded method has remarkable performance on feature selection. The results also specify how in advance that the risk level can be most accurately predicted. The study in Phase 3 proposes a risk labelling method that can continuously describe the risk during an LC event and generate LC risk profile. The risk labelling method is developed based on driving safety field theory and Dynamic Time Warping (DTW) k-means clustering. From the validation results, the LC risk profiles are recommended to be categorised into three clusters, namely, “uphill” shape, “bell” shape, and “downhill” shape. It is also found that many “uphill” LC risk profiles have higher risk, and LC risk profiles are mostly dependent on the risk between LC car and its preceding cars, which can be explained by the LC car’s purpose to overtake the preceding car in the original lane. The study in Phase 4 describes the causal relationships between key features and risky LC manoeuvre and make a comparison between risky CF and LC manoeuvres. In Phase 4, a data-driven method to develop a Copula-Bayesian Network involving multiple features is proposed. Feature inferences are conducted using the networks to illustrate the causation of the two risky manoeuvres. Several interesting findings are found in this study, for examples, the effects of surrounding environment features on the two risky manoeuvres, the relationships between crash potential and traffic density for the two manoeuvres, etc.