Research on Pedestrian Detection Algorithm in Industrial Scene Based on Improved YOLOv7‐Tiny
Ling Wang,Junxu Bai,Peng Wang,Yane Bai
DOI: https://doi.org/10.1002/tee.24075
IF: 0.923
2024-04-25
IEEJ Transactions on Electrical and Electronic Engineering
Abstract:YOLOv7 is one of the most effective algorithms for one‐stage detectors. However, when it is applied to pedestrian detection tasks in the industrial scene, it is still challenging for complex environments and multi‐scale changes of pedestrians. This paper proposes a new pedestrian detector for the industrial scene based on improved YOLOv7‐tiny and named as GP‐YOLO. First, the neck of YOLOv7‐tiny is replaced by RepGFPN structure, make full use of multi‐scale features to enhance the detection accuracy of objects with large‐scale changes. Second, a new gnconv branch is added to the feature fusion module, and the high‐order spatial interaction capability is introduced to further enhance the target detection accuracy. Finally, a lightweight method based on PModule is proposed, on this basis, a PConv bottleneck is designed to reduce the FLOPs and enhance the feature extraction. Experiments on a self‐made Industrial Pedestrian Data set show that before lightweight, the proposed algorithm achieves a 3.2% improvement in mAP@0.5:0.95 and a 3.7% improvement in Recall compared to the baseline YOLOv7‐tiny. After lightweight GP‐YOLO, compared to non‐lightweight, parameters and FLOPs are decreased by 26% and 23%, respectively, the mAP@0.5:0.95 is decreased by only 1.1% and the Recall is decreased by only 1.3%, which remains at a high level. Compared with baseline YOLOv7‐tiny, the lightweight GP‐YOLO has similar parameters and FLOPs, but the mAP@0.5:0.95 is increased by 2.1%, and the Recall is increased by 2.4%. © 2024 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.
engineering, electrical & electronic