Pedestrian trajectory prediction using goal-driven and dynamics-based deep learning framework

Abstract

Pedestrian trajectory prediction plays an important role in the design of autonomous driving systems and robotics. Recent work utilising prominent deep learning models for pedestrian motion prediction makes a limited priori assumption about human movements. Even for a few seconds, the prediction of pedestrians can be vital in avoiding road accidents for autonomous driving systems. In this paper, we present a deep learning-based framework that utilises an asymptotically stable dynamical system for pedestrian trajectory prediction. We encode goal-driven human motion in our framework via stability theory that provides expert knowledge of human motion. Our framework employs the Transformer deep learning model with an asymptotically stable dynamical system for multi-step ahead prediction in a recurrent manner. The experimental results show that our framework outperforms prominent deep learning models in pedestrian trajectory prediction on benchmark problems that include ETH/UCY and SDD datasets. Furthermore, we provide insights into the different categories of movements in the spatiotemporal nature of the pedestrian trajectory prediction. • Prediction of pedestrian trajectory can be vital in avoiding road accidents. • We present a dynamical system within a deep learning-based framework. • We use the Transformer model with the dynamical system for prediction. • Our framework outperforms related models on benchmark datasets. • We provide insights into the different categories of pedestrian trajectories.