An Advanced YOLOv5s Approach for Vehicle Detection Integrating Swin Transformer and SimAM in Dense Traffic Surveillance

In the realm of high-definition surveillance for dense traffic environments, the accurate detection and classification of vehicles remain paramount challenges, often hindered by missed detections and inaccuracies in vehicle type identification. Addressing these issues, an enhanced version of the You Only Look Once version v5s (YOLOv5s) algorithm is presented, wherein the conventional network structure is optimally modified through the partial integration of the Swin Transformer V2. This innovative approach leverages the convolutional neural networks' (CNNs) proficiency in local feature extraction alongside the Swin Transformer V2's capability in global representation capture, thereby creating a symbiotic system for improved vehicle detection. Furthermore, the introduction of the Similarity-based Attention Module (SimAM) within the CNN framework plays a pivotal role, dynamically refocusing the feature map to accentuate local features critical for accurate detection. An empirical evaluation of this augmented YOLOv5s algorithm demonstrates a significant uplift in performance metrics, evidencing an average detection precision (mAP@0.5:0.95) of 65.7%. Specifically, in the domain of vehicle category identification, a notable increase in the true positive rate by 4.48% is observed, alongside a reduction in the false negative rate by 4.11%. The culmination of these enhancements through the integration of Swin Transformer and SimAM within the YOLOv5s framework marks a substantial advancement in the precision of vehicle type recognition and reduction of target miss detection in densely populated traffic flows. The methodology's success underscores the efficacy of this integrated approach in overcoming the prevalent limitations of existing vehicle detection algorithms under complex surveillance scenarios.