Student Abnormal Behavior Recognition in Classroom Video Based on Deep Learning

For the complication and low speed of traditional human behavior recognition process, a method of student abnormal behavior recognition in classroom video based on deep learning was proposed. First of all, for the poor effect of small target identification with the original network, our method introduces an cascading improved RFB module by adding a branch to the RFB to increase the reference to the peripheral visual field. This network was called Rs-YOLOv3, and it can enhance the feature extraction capability of the original network and make full use of the shallow information to improve the identification effect of small targets. Secondly, for the character occlusion due to classroom structure and student density, the resn module of Darknet-53 in YOLOv3 was replaced with SE-Res2net module. The feeling field of each layer of the network can be increased to represent feature information in more fine-grained and realize multi-layer feature multiplexing. Finally, the border regression calculation was performed by changing the border loss function to DIoULoss. The experimental results show that the improved network SE-Res2Net-DIoU achieves 80. 1% in the accuracy of student abnormal behavior recognition, a 5. 8% improvement compared with the traditional YOLOv3, and reduces the missed recognition rate.