An improved Multi-Scale Fusion and Small Object Enhancement method for efficient pedestrian detection in dense scenes
摘要
Pedestrian detection is a fundamental task in intelligent transportation systems and autonomous driving. Deep learning methods provide effective solutions but still face problems such as variable scale of objects, small objects and pedestrian occlusion, which bring great challenges in this field. This study introduces an enhanced version of the dense pedestrian detection method, RT-DETR-MSS, which is built upon the RT-DETR framework. We make three primary contributions. First, the Multi-Scale Fusion Module is proposed, which enhances multi-scale feature fusion and captures fine-grained spatial information, improving overall performance. Second, we design the Small Object Enhance Structure, which integrates low-level fine-grained features into high-level features to improve small object detection. Third, we introduce the GSConv and Slim-Neck architecture VoV-GSCSP module. This method simplifies the network, optimizes computation, and preserves accuracy. Experiments were performed on the CrowdHuman and WiderPerson datasets to assess the effectiveness of our approach. Compared to the baseline RT-DETR model, our method improved mAP0.5 by 1.7