Driving Towards Accurate Lane Change Prediction: Insight from Machine Learning Analytics

Abstract: Anticipating lane-change patterns represents a crucial dimension within the intricate framework of lane-change decision-making, exerting a profound influence on the fluidity of traffic dynamics and the overarching spectrum of road safety. Previous studies have mostly focused on fixed sections of highways, missing the changing and complex traffic patterns that drivers experience throughout the entire highway journey. By harnessing the capabilities of a cost-effective 3-Dimentional LiDAR (Light Detection and Ranging) sensor, the instrumented vehicle platform facilitates the seamless acquisition of real-time data, particularly in challenging scenarios such as night-time conditions and obscured sightlines. Employing a precisely curated real-world dataset, christened the 'Expressway Drive Instrumented Vehicle (EDIV) Dataset,' tailored to the unique peculiarities of Indian expressway ecosystems, the study traverses the trajectory of 60 drivers across an expansive terrain spanning approximately 8000 kilometers. Beyond the mere prediction of drivers' lane-changing events, the research delves deeper into the intricate composition of lane transitions, employing a sophisticated repertoire of Machine Learning (ML) methodologies. Notably, the Extreme Gradient Boosting (XGBoost) technique emerges as the preeminent contender, showcasing unparalleled efficacy in accordance with canonical classification metrics. Culminating in the application of elucidatory Artificial Intelligence (AI) paradigms, such as SHapley Additive exPlanations (SHAP) values, to interpret the intricacies of XGBoost-derived insights, the study concludes by advocating for the seamless integration of these analytical outputs into the fabric of advanced driver assistance systems.