Pedestrian detection with Bi-Directional Feature Pyramid and Channel-Spatial Attention Modules

Abstract

In the current research, the pedestrian detection accuracy in dense scenes is low. In order to improve the detection accuracy, an improved method based on yolov5 network is proposed. The bidirectional feature pyramid network (BiFPN) is integrated into the original path aggregation network (PANet) to enhance its capability. This integration aims to bolster the fusion of multi-scale features, consequently amplifying the detection proficiency for pedestrian targets. In order to retain more feature information and improve the feature extraction ability of the backbone network, the residual structure VBlock is added. This method integrates the CBAM attention mechanism. It enables a more comprehensive expression of information within the feature map. The proposed algorithm is evaluated using the CrowdHuman dataset for training and testing. Experimental results indicate improvements of 1.8% in accuracy, 2.4% in recall, and 2.5% in average accuracy compared to the original network. This confirms the algorithm's effectiveness in enhancing human target detection precision in crowded scenes.