Research on YOLOv3 pedestrian detection algorithm based on channel attention mechanism

This paper presents YOLOv3 pedestrian identification algorithm based on the channel attention mechanism for addressing the concerns of poor detection accuracy and low placement accuracy of standard pedestrian target recognition methods. To begin with, the structure of the feature extraction network Darknet-53 is enhanced, and the original residual structure is adjusted using the channel attention module ECA module to increase feature extraction capabilities. Secondly, using the spatial pyramid pooling module of three different receptive fields to fuse multi-scale features, thereby inputting features of different sizes, so that the network can learn the target features more comprehensively. Finally, the k-means++ method is used to the Caltech pedestrian data set to perform clustering analysis in order to find a more appropriate prior frame. The experimental results demonstrate that, when compared to the YOLOv3 model, the accuracy of the proposed enhanced YOLOv3 model is raised by 2.4 percent, and the recall rate is increased by 1.9 percent under the assumption of a minor increase in the quantity of model calculation.