A Collaborative YOLOv5-GA-Net Framework for Real-Time Stereo 3D Object Detection in Autonomous Driving

Accurate and real-time 3D perception is essential for autonomous driving, particularly under occlusion, low illumination, and adverse weather conditions. To enhance the robustness and efficiency of stereo-based 3D detection, this study proposes YGS-Net, a dual-branch collaborative framework that integrates an improved YOLOv5 detector with a guided aggregation stereo-matching network. The detection branch incorporates coordinate attention, a bidirectional feature pyramid, and a DIoU regression loss to strengthen multi-scale feature representation and box localization. The depth branch produces dense and reliable disparity maps through GA-Net. A collaborative fusion mechanism maps each 2D detection to its corresponding disparity region and estimates depth via region-averaged disparity. Experiments on the KITTI dataset demonstrate that YGS-Net achieves 75.13% AP_3D (moderate) for the "car" category, outperforming PL++, DSGN, LIGA-Stereo, and YOLOStereo3D. Real-world tests show a relative ranging error below 2.5% within 20 meters and real-time performance at ~ 42 FPS. The results indicate that YGS-Net provides a balanced and reliable purely visual 3D perception solution for complex traffic environments.