A Keypoint-guided Pipeline for Safety Violation Identification
Guoli He, Donglian Qi · 2020
In this paper, a keypoint-guided pipeline for safety violation identification is proposed. Firstly, a top-down keypoint detector is adopted to detect human areas and 17 keypoints in each human bounding box. Secondly, related regions such as head, trunk, etc. are accurately located. Finally, these regions are cropped from the original image and fed into a unified classifier for feature extraction and violation identification. The experimental results show the superiority of the proposed pipeline than other existing counterparts. An average accuracy of 91.2% is achieved on identifications of not wearing helmet, work clothes and smoking.