Enhancing trajectory prediction of simultaneous collision avoidance and interaction modelling through parameter learning using machine learning

Abstract

Autonomous driving is a hot topic throughout the world at present. When it comes to autonomous driving, tracking of road agents like vehicles and pedestrians is always an important issue to consider as it plays a vital role in trajectory prediction. Trajectories of road agents are dominated by various dynamic constraints. Simultaneous Collision Avoidance and Interaction Modelling (SimCAI) is a novel motion model for predicting the motion of road agents. It consists of techniques to avoid collisions between the road agents and to model safety interactions effectively. This research aimed to enhance trajectory prediction in traffic videos by optimizing parameter values and integrating the SimCAI model with a framework called TrackNPred for comparison purposes. Three machine learning search algorithms were employed to explore the parameter space and identify optimal configurations. Integration of SimCAI with TrackNPred allowed for evaluating and comparing its performance with Reciprocal Velocity Obstacles (RVO) in realworld scenarios. The research focused on improving SimCAI by learning the context-dependent parameter values through the search algorithms. Additionally, comparison techniques were employed to evaluate the accuracy of trajectory predictions. By identifying the best parameter values through the search algorithms, a reduction in Binary Cross-Entropy (BCE) loss was observed, indicating improved performance.