Combine Object Detection with Skeleton-Based Action Recognition to Detect Smoking Behavior
Xuanrui Zhang, Xieyang Su, Junbo Yu, Weihong Jiang, Shengchun Wang, Yuan Zhang, Zhiyong Zhang, Liang Wang · 2021
This paper detects violations of smoking in non-smoking areas by construction workers, uses the YOLO object detection algorithm combined with the Kalman filter to track the human body, then uses Alphapose's human pose estimation to obtain the key points of the human body. We input the key points into Spatial-Temporal Graph Convolutional Networks for preliminary identification of workers’ smoking behavior. However, this will lose the texture feature information of the image, resulting in drinking, scratching, etc. will also be recognized as smoking. Therefore, based on action recognition, the YOLO object detection algorithm is added for the second time to extract the image texture of the workers’ smoking based on a face-gesture-cigarette pattern and achieve the joint determination of posture and texture in order to improve the detection accuracy and robustness of smoking recognition.