Research on deep learning-based instruction action recognition method for assistant duty officers
Haonan Liu, Wenzhen Kuang · 2023
With the vigorous development of railway industry, railway passenger transport and freight has become the most important link in today's transport industry. The operation of receiving and sending trains at the station is related to the transportation safety of the whole railway, so the action standard of the railway assistant on duty is more strict. In this paper, an improved action recognition algorithm for assistant duty officers based on dual flow convolutional neural network embedded attention mechanism is proposed. First, the TV-L1 method was used for optical flow processing of the video, and the mediapipe was used to obtain the coordinate data of human bone points. The two kinds of data were used as network input for training. Then, the trained network model is fused, and the two flow convolutional neural networks embed attention mechanisms. VGG16 network is selected for both spatial network and time network. Finally, the two models were weighted and fused by softmax function as the output result to complete the command action recognition of the assistant on duty. Experiments were carried out on the dataset UCF101, and the accuracy of the model on this dataset was 86.7%. Experimental results show that the improved network model has better recognition effect.