Pedestrian Intention Prediction Using Uncertainty-Aware Human Pose Estimation

Accurate pedestrian intention prediction is vital for autonomous vehicle safety. Human pose keypoints offer valuable cues, but uncertainty in their detection limits reliability. This paper introduces an uncertainty-aware prediction framework. We utilize multiple keypoint detectors (ViTPose variants) concurrently and quantify uncertainty via the variance in their predictions. This uncertainty measure is integrated as an input feature into an LSTM-based intention prediction model. Evaluating on the PIE dataset, our uncertainty-aware approach achieved 85% accuracy and 86% F1-Score, improving significantly over a baseline LSTM without uncertainty (+8% accuracy, +9% F1Score). Results demonstrate that explicitly modeling keypoint uncertainty enhances the precision and robustness of pedestrian intention predictions for AVs, achieving performance comparable to more complex models while focusing on keypoint data.