Student classroom behavior detection method based on SCB-DETR

To address the problems existing in the real classroom environment, such as complex backgrounds, mutual occlusion among personnel, and similar behavioral characteristics, this paper constructs a student classroom behavior detection model based on SCB-DETR. By designing the CSMF multi-semantic collaborative fusion module, the ability of the backbone network to capture and mine detailed information such as edge textures, relative positions, and high-level representative features is strengthened. In the feature fusion stage, the SFS_BiFPN feature fusion network is proposed. This network integrates the SFS feature screening and BiFPN methods and introduces the DySample upsampling operator, thereby enhancing the fusion effect of multi-scale features. The WIoU-v3 is integrated with Facaler-IoU to optimize the loss function and alleviate the sample imbalance phenomenon during the algorithm training process. Experiments were conducted on the self-made dataset CSMBD-Dataset and the public dataset SCB-Dataset3, and the results confirmed that the proposed model has better detection performance in the classroom scenario.