<p>Ablation Experiments.</p>

To address the problem of insufficient detection accuracy for dense targets, small targets and partially occluded objects in complex road scenarios, an improved object detection model, SSH-YOLO, is proposed. On the basis of YOLOv8n, the model optimizes and improves performance through a three-level collaborative architecture: 1) introduce the spatial and deep conversion (SPDConv) module in the backbone network to replace the traditional step downsampling with nonstep convolution, retain the fine-grained features of small targets, and solve the feature loss problem of low-resolution images; 2) embed the spatial and channel collaborative attention module (SCSA), through cross-scale feature fusion (SMSA) and channel weight progressive optimization (PCSA), focus on the key visible areas of the occluded target and suppress background interference such as roadside vegetation; and 3) add a new 160 × 160 resolution small object detection head, combined with the original P3‒P5 layer to form a four-level detection system, covering long-distance small targets < 32 × 32 pixels. The experimental results show that the improved model performs well on the self-built RoadScene-Complex dataset and four public datasets BDD100K: 0.729 (12.4% higher than YOLOv8n) on the RoadScene-Complex dataset mAP@0.5 (12.4% higher than YOLOv8n) and 0.868 (7.6% higher than the KITTI dataset) mAP@0.5). COCO small target subset mAP@0.5 to 0.585 (up 16.5%), and CityPersons occluded scene mAP@0.5 to 0.739 (up 22.8%). At the same time, it maintains lightweight characteristics and has an inference speed of up to 60 FPS to meet the needs of real-time on-board detection. The research results provide a balanced solution of “accuracy-speed-lightweight” for high-precision target detection in complex traffic scenarios, especially in small target and occlusion scenarios.