A phased behavior recognition and detection algorithm based on multi-network fusion

Aiming at the problems of small detection range and low recognition accuracy in the existing examinee behavior recognition in the examination room, a multi network fusion algorithm for examinee behavior recognition is proposed. The lightweight detection network Yolov4 Tiny is selected and improved for candidate positioning. First, channel space dual attention mechanism CBAM is embedded in the trunk, which solves the problem of difficult identification of small targets and occluded targets in the examination room. Secondly, the introduction of PPM pyramid pool structure after feature extraction can improve the network's ability to obtain global information. Then the improved network is integrated into the Alphapose human posture estimation model to extract the coordinate information of the examinee's skeleton key points, and finally the behavior classification is carried out through the space-time map convolution neural network ST-GCN. The experiment shows that the pre training model is obtained in the data set NTU-RGB+D by means of transfer learning, and the average accuracy rate of four types of behavior recognition on the examinee's behavior data set finally reaches 94.6%, which can effectively complete the examinee's behavior recognition and detection task in the examination room.