A Robust Substation Indoor Personnel Action Recognition Network and Online Framework

It is crucial for operation safety that substation staff work indoors as required. Therefore, it is necessary to monitor the behavior of workers to reduce the occurrence of dangerous practices. Work in substations often involves interaction with equipment, leading to frequent occlusions that challenge the robustness of action recognition models. Additionally, monitor typically needs to be performed online, requiring the extraction of action segments from the live video stream for label classification. We have established a dataset specifically for the recognition of certain actions about indoor personnel in substations. A special processing method for skeletons is adopted and the recognition network is improved, which surpasses the accuracy of conventional Graph Convolutional Networks and significantly outperforms them in terms of robustness. Furthermore, we have proposed three online action detection frameworks according to the actual recognition scenarios in substations, balancing both inference accuracy and efficiency. Experiments conducted on our self dataset have verified the effectiveness of the proposed methods.