Open AccessYOLOv5s-FCG : An Improved YOLOv5 Method for Inspecting Riders’ Helmet Wearing
Author(s) -
Pengfei Wang,
Hanming Huang,
Mengqi Wang,
Bingjun Li
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/2024/1/012059
Subject(s) - bottleneck , adaptability , feature (linguistics) , computation , computer science , artificial intelligence , generalization , volume (thermodynamics) , scale (ratio) , data set , engineering , computer vision , mathematics , algorithm , embedded system , geography , ecology , mathematical analysis , linguistics , philosophy , physics , cartography , quantum mechanics , biology
Manual inspection of riders’ helmets is time-consuming and labor-intensive, low in efficiency and small in coverage. Aiming at this shortcoming, this paper proposed an improved YOLOv5s-FCG(FourLayers, CBAM attention, GhostBottleneck) helmet wearing detection method based on YOLOv5 (You Only Look Once). Based on the smallest volume of YOLOv5s in YOLOv5 series, the network was improved, the shallow feature detection layer was added, the three-scale feature detection was changed to four-scale feature detection, and the up sampling was increased by four times. Add the CBAM attention module; Use lightweight GhostBottleneck instead of Bottleneck structures. The results in our experiments show that YOLOv5s-FCG raises the average detection accuracy (mAP) by 2.0% compared with YOLOv5s on the ourselves’ riding safety helmet data set and 1.5% on the NWPU-VHR-10 public data set. The proposed algorithm not only ensured the detection rate, volume, computation and number of parameters, but also improved the detection accuracy. And it had good adaptability and generalization ability in complex road environments such as poor light and small targets.