DA-SlowFast Network for Abnormal Action Detection in the Workshop

The detection of abnormal workshop actions plays a vital role in industrial production, which gives timely warning when the workshop employees behave abnormally and thus greatly reduce the probability of workshop safety accidents. Despite the tremendous advances in action recognition and detection technology in recent years, real-time abnormal action detection remains a challenging task due to the complexity of video data and the huge computational costs. In order to achieve high accuracy detection of abnormal actions in complex workshop scenarios, this paper proposes a DA-SlowFast algorithm for workshop abnormal action detection. The algorithm can be used to detect six common types of abnormal actions in the workshop (talking on the phone, sleeping on the workbench, falling, running violently, fighting, playing with cell phones). Based on the original SlowFast network, we incorporate the hybrid domain attention mechanism (DANet) in the Slow branching and lateral connectivity processes, in order to improve the model's detection accuracy for static actions and its ability to capture fine-grained motion details. In addition, for purpose of solving the problem of unbalanced samples of various types of actions in the video dataset, we replace the Cross Entropy Loss function for the action classification task in the original model with the Focal Loss function. The results of the ablation experiments show that the efficient hybrid attention mechanism introduced in the SlowFast model significantly improves the detection accuracy. And the final DA-SlowFast model also obtains the highest detection accuracy compared to other similar models on the workshop dataset.