It is challenging to precisely predict pedestrian crossing intentions in autonomous driving, which requires combining changing traffic scene with dynamic pedestrian behaviors. Past studies have always resorted to pedestrians' historical behaviors to forecast their crossing intentions. However, pedestrians' crossing intentions are more related to their future behaviors. Moreover, deep learning-based crossing intention prediction lacks interpretability, limiting the practical application in intelligent vehicles. To tackle the mentioned issues, we propose a Multi-Task Mutual Guidance Network (MTMGN), which optimizes the primary task of pedestrian crossing intention prediction by introducing an auxiliary task of pedestrian posture generation. The crossing intention prediction task consists of two modules: (1) A multi-source input fusion module is proposed to extract context-constraint features based on multiple inputs, including historical pedestrian postures, semantic segmentation maps, and pedestrian-vehicle distances. (2) A future posture injection module is proposed to forecast whether the pedestrian will cross based on context-constraint features and generated future postures. The pedestrian posture generation task leverages the conditional diffusion model to generate future pedestrian postures that adhere to the traffic scene based on the extracted context-constraint features. MTMGN is evaluated on two public datasets, JAAD and PIE, with F1/Recall values of 73%/89% and 92%/90%, respectively. Both qualitative and quantitative evaluations indicate that MTMGN can precisely anticipate pedestrians' crossing intentions, benefitting from the generated pedestrian postures. Moreover, the generated postures provide intuitionistic interpretability for crossing intention prediction.
Explainable Pedestrian Crossing Intention Prediction Based on Multi-Task Mutual Guidance Network
IEEE Transactions on Intelligent Vehicles ; 10 , 2 ; 1342-1352
2025-02-01
2693292 byte
Article (Journal)
Electronic Resource
English