Optimized deep learning for pedestrian safety in autonomous vehicles

This study focuses on enhancing pedestrian detection for autonomous driving and intelligent surveillance systems, where challenges like complex backgrounds, obstructions, and small target sizes can hinder accuracy. The researchers optimized the YOLOv8 model by redesigning its neck structure using the BiFPN (Bidirectional Feature Pyramid Network), reducing parameters, size, and computational load. They also integrated Coordinate Attention into the SPPF (Spatial Pyramid Pooling-Fast) layer for improved localization and feature integration. Additionally, the CIoU loss function was applied to refine anchor regression predictions for better edge positioning accuracy. Experimental evaluations on the KITTI, Caltech Pedestrian, and CityPersons datasets demonstrate that PD-YOLOv8 achieves superior detection performance compared to state-of-the-art methods. The findings underscore the model’s robustness across diverse environmental conditions, highlighting its potential for real-world deployment in autonomous vehicle perception and intelligent surveillance applications.