Due to technological advances in the automotive field, advanced driver assistance systems have attracted increasing interest from various research and development entities. Predicting road users' future trajectories remains an active research challenge for advanced driver assistance systems. Accurate Trajectory Prediction (TP) allows anticipation of surrounding road users' future motion, enabling timely safety-critical interventions such as speed regulation and emergency braking in unexpected driving situations. Recent advances in TP methods based on artificial intelligence have demonstrated remarkably accurate results compared to traditional methods. However, many of these models require a high computational burden, which makes their deployment on embedded architectures with constrained resources challenging. To overcome these constraints, TP models need to be lightweight and efficient to meet the real-time and power consumption requirements of advanced driver assistance systems. In other words, they must maintain high accuracy while guaranteeing low computational load and rapid inference. This paper presents a comparative and experimental review of state-of-the-art vehicle TP models. First, we propose a new taxonomy based on the operating environment, the trajectory output type, and the employed modeling approach to classify existing methods. Then, we evaluate representative approaches w.r.t the taxonomy in terms of accuracy, model complexity, computational performance, and real-time feasibility across a high-performance architecture and an embedded architecture. Finally, we discuss the evaluation results and present key conclusions and future directions.
Skalli et al. (Thu,) studied this question.