YOLO-TAD: a distribution-aware multi-scale detection network for traffic targets in complex scenarios

Abstract In real-world traffic environments, complex factors often cause small targets to be missed or occluded targets to be misidentified during vehicle and pedestrian detection, making this task a challenging problem in the field of object detection. To address the challenges of visual perception in complex traffic scenarios, this paper redefines vehicle and pedestrian detection as a geometric measurement problem based on object detection and proposes YOLO-TAD (TriResMSConv, axial rotation spatial attention (ARSA), and D 2 Head), an intelligent perception and detection network designed to enhance measurement accuracy and spatiotemporal consistency. This improved network incorporates three key innovations: TriResMSConv (triple residual multi-scale segmentation convolution), ARSA, and a fully decoupled D 2 Head (densely distributed perception detection head) based on distributed regression. TriResMSConv enhances feature representation through multi-scale residual branches; ARSA performs axial decomposition in the spatial dimension and applies (rotation position encoding) to emphasize orientation dependence; and the D 2 Head integrates IoU perception with distributed regression mechanisms to improve localization accuracy. Additionally, this method employs full projection IoU (CPIoU) to increase the accuracy and stability of bounding box recognition. Experimental results demonstrate that the proposed method significantly improves detection performance on the SODA-10M autonomous driving perception dataset. Compared with YOLO11s, this model achieves improvements of 3.0% and 2.2% in mAP@50 and mAP@50–95, respectively. Subsequent generalization experiments conducted on the KITTI and BDD100K datasets confirmed this finding.