Unified Deep Learning for Real-Time Pedestrian Detection, Pose Estimation, and Tracking

This study introduces a novel unified deep learning framework for real-time pedestrian and Vulnerable Road User (VRU) detection, pose estimation, and tracking using YOLOv8. Unlike traditional approaches that separately handle these tasks, our integrated multi-task model leverages YOLOv8’s advanced multi-scale feature extraction and optimized architecture to efficiently perform simultaneous detection, pose estimation, and tracking. Experimental evaluations demonstrate superior performance compared to baseline YOLOv8 configurations, achieving an [email protected] of 57.2%, OKS of 76.1% (COCO dataset), MOTA of 67.1%, and IDF1 of 64.3%. The framework's robust performance is validated through comprehensive testing under realistic urban scenarios and challenging conditions. By effectively addressing limitations in current autonomous vehicle (AV) perception systems, such as handling occlusions, varying lighting, and dense pedestrian environments, this integrated approach significantly enhances AV safety and navigation reliability at critical junctions and pedestrian crossings.