Research and application of Faster RCNN-based video action detection model

This paper illustrates the research process of a Faster RCNN-based video action detection model and its application to assist safety production in the power industry. The model combines the Faster R-CNN method and the dual channel theory to achieve all the use of deep learning techniques in obtaining the area of interest and extracting video characteristics. Using the dual-channel network model extraction feature, two different Faster R-CNN networks are used to extract apparent features and action features from the video, and then the two types of features fused to form a time-space domain feature, used to represent the original features of the video. Finally, when the extracted feature is classified, the frame and the timing level are considered, and the action pipeline model is constructed when the video recognition task is studied. The model can sequentially connect the interest area on the frame and form a campaign in accordance with the frame order and selecting the optimal campaign pipe as a video final action judgment result. The category of the optimal campaign pipe is marked as a classification of the entire video operation, which constitutes an interested region set of the optimal campaign pipe as a position of the entire video operation. The algorithm model trained with this model performs well in practical power industry application scenarios, assisting supervisors in timely detection of potential hazardous situations in the form of irregularities in personnel operations, equipment and environmental abnormalities, and assisting in improving the operational efficiency and safety management level of power enterprises.