Pedestrian Intention Prediction

Pedestrian intention prediction estimates a person’s likely near-future action—especially crossing versus waiting—and often forecasts the trajectory as well, helping autonomous vehicles and robots choose safe responses before motion is clear. Methods combine visual appearance and motion with pose, gaze, vehicle state, and scene or social interactions, using temporal transformers and multimodal fusion alongside structured probabilistic models and multimodal trajectory predictors such as conditional variational autoencoders. Vision-language models are also being adapted to answer intent questions from video, while event-camera and language-guided planning approaches extend the task to low-latency sensing and downstream control; evaluation commonly uses driving datasets such as PIE and JAAD. Strong benchmark results coexist with open challenges in uncertainty and out-of-distribution robustness, cross-dataset transfer, interpretable explanations, and reliable prediction under occlusion or ambiguous behavior.