CD-PIP: Conditional Diffusion for Pedestrian Intention Prediction under Occlusion

Abstract

Pedestrian crossing intention prediction is a core perceptual capability for intelligent vehicles operating in urban traffic environments. Most existing intention-prediction methods assume that pedestrian motion observations are complete, whereas in real traffic scenes, vehicles, other pedestrians, and roadside infrastructure often cause short-term occlusions, leading to missing motion cues such as bounding boxes, center coordinates, and ego-vehicle speed. Reducing the impact of such missing cues on intention prediction requires not only effectively recovering the missing motion information, but also ensuring that the recovered results genuinely serve the downstream prediction task. To this end, this paper proposes CD-PIP, an occlusion-robust pedestrian intention prediction framework for occluded scenes. The framework builds explicit conditions from visible observations, occlusion masks, and diffusion time steps, learns the conditional distribution of pedestrian motion features in the missing regions, and jointly optimizes a reconstruction loss together with an intention classification loss, so that the imputed results balance reconstruction accuracy against downstream discriminative requirements. Experiments with three random seeds on the PIE dataset show that conditional diffusion imputation reduces the masked MAE from 0.0419 to 0.0109, a relative decrease of 73.9%, while raising the F1 score from 0.8960 to 0.9053.