Research on Industrial Human Action Recognition Based on Improved SlowFast

In industrial site, safety accidents caused by unsafe actions of personnel are unfortunately common. Swift and accurate recognition of human actions can timely alert on-site personnel and reduce the risk of injuries and property damage caused by production accidents. This paper proposes an improved action recognition algorithm based on the Slow Fast framework. It introduces the C3DS attention module and the MH-C3DS multi-head attention module to enhance the network's capability in extracting different action features. The original single-classifier structure is transformed into a multi-branch structure with independent classifiers to improve the recognition accuracy for multiple actions. A dataset of human actions in industrial site is constructed to analyze the algorithm's recognition performance in real-world industrial site. With the incorporation of attention modules and the multi-classifier structure, the recognition accuracy of the algorithm is improved by 9.8%, achieving an overall accuracy of 77.4% for the improved recognition algorithm. The experimental results demonstrate that the improved recognition algorithm performs well in real industrial site, validating the effectiveness of the proposed algorithm enhancements.