ECTR-YOLOv5:Pedestrian detection in dense scenes based on improved YOLOv5

Abstract

Abstract Pedestrian detection technology has reached a relatively mature level in sparse environments. However, accurate pedestrian detection in packed scenes still presents challenges owing to factors such as occlusion, high crowd density, and scale changes. This study proposes a pedestrian detection algorithm for dense scenes based on the YOLOv5 model and several key modifications have been introduced to enhance performance, First, the backbone network incorporates the new attention mechanism network module. This addition improves pedestrian detection accuracy across multi-scale changes by effectively integrating both channel and spatial information. Second, the connection of the P2 layer is implemented to enhance the capture of features at different scales. This is particularly beneficial in reducing the missed detection rate of distant pedestrians. Third, using a Weighted Bidirectional Feature Pyramid Network neck network not only reduces model complexity in neck networks but also efficiently and quickly achieves multi-scale fusion. Finally, the prediction network is augmented with the adaptive spatial feature fusion block (ASFF). This integration enhances multi-scale feature maps, accommodating spatial variation in prediction uncertainty. Experimental results demonstrate that the optimized YOLOv5(ECTR-YOLOv5) achieves performance improvements across both the WiderPerson and CrowdHuman datasets. These enhancements showcase excellent performance and robustness under varying environmental conditions, rendering the algorithm more suitable for real-time target detection applications. In summary, the proposed modifications to the YOLOv5 model, as embodied in ECTR-YOLOv5, represent a significant advancement in addressing the challenges of accurate pedestrian detection in packed scenes. The algorithm's improved performance and robustness make it well-suited for deployment in real-world scenarios where detection is in densely populated environments.