Pedestrian detection and tracking using an enhanced YOLOv9 model for automotive vehicles

Abstract

Pedestrian detection in autonomous driving systems is challenging due to complex urban environments, where pedestrians often blend with surrounding objects, affecting detection accuracy. To address these challenges, this paper presents a novel multi-object tracking (MOT) model combining the YOLOv9 detection algorithm with DeepSORT tracking. Key improvements include replacing the Backbone’s RepNSCPELAN4 module with a CAM context enhancement module for better feature extraction from small or occluded pedestrians, integrating the AFF channel attention mechanism to resolve semantic and scale inconsistencies, and introducing the AKConv dynamic convolution for enhanced contextual information capture in dynamic scenes. We propose evaluating a public benchmark that integrates three datasets: KITTI, EuroCity, and BDD100K. The improved YOLOv9-DeepSORT model shows strong performance across different datasets. On the KITTI Dataset, the model achieved 98.12 % precision, 92.48 % recall, a [email protected] of 95.73 %, and a [email protected]:0.95 of 90.68 %. Meanwhile, on the EuroCity Persons dataset, the results were 95.12 % precision, 90.55 % recall, a [email protected] of 94.50 %, and a [email protected]:0.95 of 79.12 %. These results highlight the model’s effectiveness in different pedestrian detection and tracking scenarios, demonstrating improved performance in both urban and challenging environments.