Faster pedestrian crossing intention prediction based on efficient fusion of diverse intention influencing factors

Predicting pedestrian crossing intention to ensure pedestrian safety has garnered significant attention in autonomous driving. Balancing the accuracy and real-time performance of prediction is a challenging endeavor. This paper introduces Faster-PCPNet, a rapid pedestrian crossing intention prediction network that jointly predicts intention based on pedestrian pose, ego-vehicle speed, pedestrian bounding box, and the novel triple quasi-polar coordinate. Faster-PCPNet incorporates the Temporal-Channel sharing enhanced topology Graph Convolution (TCGC) module designed in this paper, allowing for in-depth exploration of significant pedestrian crossing action features. This paper innovatively establishes a quasi-polar coordinate system, addressing the challenge of efficiently representing pedestrian crossing intention during relative motion between pedestrians and vehicles. Faster-PCPNet achieves accuracy rates of 89% on the JAAD dataset and 94% on the PIE public dataset. Experimental results highlight the real-time performance and reliability of the proposed model on mobile devices in traffic scenarios, effectively mitigating traffic accidents caused by pedestrian crossing. Our code and model are available at https://github.com/zjrcczu/Faster-PCPNet.