Pedestrian-Vehicle Information Modulation for Pedestrian Crossing Intention Prediction

Pedestrian crossing intention prediction (PCIP) is crucial for pedestrians' safety in autonomous driving. Existing methods do not use the interaction between pedestrians and cars for their prediction. In this paper, we argue that pedestrians' intentions are highly dependent on their interaction with the environment. Specifically, the trajectories of pedestrians and the dynamic of vehicles jointly affect the entire traffic environment in the future. Therefore, in this paper, we propose a novel pedestrian-vehicle information modulation network (PVIM). Particularly, we first propose a pedestrian-vehicle spatial context (PVSC) that effectively models the spatial dynamics between the pedestrian and ego-vehicle. Second, we design a temporal bilinear attention module that removes temporal redundancy and consolidates temporal correlation for more accurate predictions. We have conducted extensive experiments on the PIE pedestrian action prediction benchmark and have achieved state-of-the-art performance. Specifically, the proposed method achieves an accuracy of 0.91, outperforming the previous best by 2%. The code will be on https://github.com/icecherylXuli/PVIM.