A Video Detection Method of Smoking Behavior Based on Key Point a Priori and DN DETR
Di Yan, Hongyi Wang, Xinjun Zhu · 2023
With the development of artificial intelligence and computer vision technologies, while Transformer transitions from the field of natural language processing to the field of object detection, video detection of smoking behavior based on DETR and its variant models has received widespread attention. However, the computational complexity of Transformer is too high, which leads to serious hardware consumption and low detection efficiency. In this work, we propose a smoking behavior detection method that combines the human key point detection algorithm BlazePose and DN DETR, which improves the detection efficiency by clearing redundant frames in surveillance videos and arrived detection accuracy of 94% for$AP_{50}$. BlazePose ensures that the video frames transmitted to DN DETR contain people. And DN DETR is used for the high accuracy detection of smoking behavior in the video frames. The experimental results have shown that the proposed method is of high detection efficiency for smoking behavior and good practicality in engineering applications.