Recognition of Unsafe Behavior of Electrical Workers based on Spatial-Temporal Visual Attention
Jun W. Wu, Guohui Li, Jingwen Zhang, Fuxing Wang, Qin Huang · 2023
It is helpful for accident prevention to recognize unsafe behavior of electrical workers through computer vision. Surveillance video contains a lot of temporal redundant information, which affects the accuracy of unsafe behavior detection. To this end, this paper proposes a visual attention-based keyframe extraction method that eliminates temporal redundancy, reconstructs surveillance video, and uses it for unsafe behavior recognition. First, the spatial attention model is used to extract a saliency map for each frame. Based on the pixel-level features of the saliency map, the temporal attention model establishes the temporal connection between all frames and obtains the importance degree of each frame. The importance degree of the frame is ranked from high to low. An optimization functional model is established that combines the similarity degree and importance degree between frames to select the corresponding keyframe, based on which the new video is reconstructed. Finally the new video is fed into a traditional behavior recognition algorithm to recognize unsafe behavior. The experimental results show that the keyframe extraction method based on spatial-temporal visual attention can effectively improve the accuracy of unsafe behavior recognition, which is increased from 80.75% of the traditional method to 86%.