Research on Vehicle Detection Algorithms Based on Deep Learning in Complex Environments

This study proposes an enhanced vehicle detection model with three key improvements to address YOLOv8’s limitations in complex traffic scenes, particularly its poor handling of scale variation, which leads to missed detections and localization errors. First, the SFSCConv module uses a multi-branch structure to capture multi-scale features. Second, the EUSC module integrates channel attention into upsampling to preserve fine details. Third, Inner Shape-IoU leverages segmentation masks to align the IoU calculation with irregular object boundaries for better localization. On the KITTI dataset, the model achieves an mAP of 88.6% for small, occluded, and irregular objects , a 6.5% improvement over YOLOv8s, and an overall mAP of 85.42%, a 6.9% improvement. The model also demonstrates strong generalization on the BDD100K dataset. These advances offer a more robust solution for autonomous driving and intelligent traffic systems.