Confusion matrix and minimum cross-entropy metrics based motion recognition system in the classroom

This research proposes a motion recognition system for early detection of students' physical aggressive behavior in the classroom. The motion recognition system recognizes physical attacks so that teachers can resolve disputes early to reduce other greater injuries. In the beginning, cameras were used in this system to monitor students’ classroom activities and to obtain body images by removing background and saliency maps. Two angles from arm to shoulder and shoulder to the center of the body are then measured and the velocity between the two frames from the movement of the body is detected, and use these angle and velocity values as the criterion for judging whether it is a physical attack. In the end, the accuracy of the proposed algorithms is verified by using the confusion matrix based on machine learning and the minimum cross entropy based on neural networks. The simulation proves that the proposed algorithm can correctly detect the attack behavior of the collected videos.