Improved YOLO-based Pedestrian Detection Algorithm
Cong Chen, Huasheng Zhu, Qiaofeng Ren, Rui Luo, Hong Lan · 2023
When the YOLO algorithm is directly used for pedestrian detection, it suffers from the problem that the algorithm is complex and not suitable for deployment in mobile terminal. Therefore, an improved YOLO-based pedestrian detection algorithm is proposed. The algorithm is based on YOLOv5n by introducing CBAM attention mechanism, which makes the network pay more attention to the feature information of channel and space, and is conducive to extracting the important features of the target. Secondly, the lightweight CARAFE module is selected for the up-sampling module on the neck layer, which can reduce the complexity of the algorithm and obtain a larger perceptual field to improve the learning ability of the network. Finally, the Focal-EIoU loss function is chosen to solve the problem of inaccurate bounding box regression due to variable human morphology in pedestrian detection tasks. Experimental results show that the mAP value of this algorithm on INRIA Person Dataset reaches 98.4%, which is 2.4% higher than the original algorithm. The size of the improved algorithm model is less than 4 MB, which improves the detection performance while maintaining a small scale, and is more conducive to deployment on mobile terminal.