Bao Wentao ; Yu Qi ; Kong Yu

Abstract

Traffic accident anticipation aims to predict accidents from dashcam videosas early as possible, which is critical to safety-guaranteed self-drivingsystems. With cluttered traffic scenes and limited visual cues, it is of greatchallenge to predict how long there will be an accident from early observedframes. Most existing approaches are developed to learn features ofaccident-relevant agents for accident anticipation, while ignoring the featuresof their spatial and temporal relations. Besides, current deterministic deepneural networks could be overconfident in false predictions, leading to highrisk of traffic accidents caused by self-driving systems. In this paper, wepropose an uncertainty-based accident anticipation model with spatio-temporalrelational learning. It sequentially predicts the probability of trafficaccident occurrence with dashcam videos. Specifically, we propose to takeadvantage of graph convolution and recurrent networks for relational featurelearning, and leverage Bayesian neural networks to address the intrinsicvariability of latent relational representations. The derived uncertainty-basedranking loss is found to significantly boost model performance by improving thequality of relational features. In addition, we collect a new Car Crash Dataset(CCD) for traffic accident anticipation which contains environmental attributesand accident reasons annotations. Experimental results on both public and thenewly-compiled datasets show state-of-the-art performance of our model. Ourcode and CCD dataset are available at https://github.com/Cogito2012/UString.