VOSTN: Variational One-shot Transformer Network for Pedestrian Trajectory Prediction

Abstract

The accurate and reliable prediction of pedestrian future trajectories is of crucial significance for ensuring the safe navigation of autonomous driving systems. This paper introduces a novel approach called the Variational One-Shot Transformer Network (VOSTN) for the prediction of future trajectories within the 2D on-board domain. VOSTN presents an innovative method called the One-Shot Generation Block, which aims to generate queries simultaneously in order to predict future trajectories in a parallel manner. The utilization of the parallel prediction effectively addresses the issue of error accumulation resulting from autoregression, improves the efficiency of inference time, and enhances the precision of long-term trajectory prediction. And the cross-attention module investigates the inter-relationship between trajectory and ego-motion. Given the inherent stochasticity of pedestrian movement, we employ the Conditional Variational AutoEncoder to forecast available multimodal trajectories. Experimental results demonstrate that our model effectively exploits the information associated with trajectory and ego-motion, leading to the acquisition of more comprehensive feature representations. Moreover, our model outperforms the performance of existing methods on the two 2D on-board domain datasets. Our deterministic/multimodal prediction models show a reduction in the bounding box center final displacement error by 8% / 9% and 0.7% / 3% on PIE and JAAD, respectively, when compared to the most optimal baseline.