Predicting Pedestrian Intentions in Self-Driving Cars: Leveraging Non-Visual Features and Semantic Mapping

Predicting pedestrians' intentions to cross paths with cars, particularly at intersections and crosswalks, is critical for autonomous systems. While recent studies have showcased the effectiveness of deep learning models based on computer vision in this domain, current models often lack the requisite confidence for integration into autonomous systems, leaving several unresolved issues. One of the fundamental challenges in autonomous systems is accurately predicting whether pedestrians intend to cross the path of a self-driving car. Our proposed model addresses this challenge by employing convolutional neural networks to predict pedestrian crossing intentions based on non-visual input data, including body pose, car velocity, and pedestrian bounding box, across sequential video frames. By logically arranging non-visual features in a 2D matrix format and utilizing an RGB semantic map to aid in comprehending and distinguishing fused features, our model achieves improved accuracy in pedestrian crossing intention prediction compared to previous approaches. Evaluation against the criteria of the JAAD database for pedestrian crossing intention prediction demonstrates significant enhancements over prior studies.