CPID-YOLOv8: Student Behavior Detection Algorithm Based on Improved YOLOv8

Abstract

To address challenges such as dense small targets, occlusions, and multi-scale spanning in natural classroom teaching scenarios, an improved YOLOv8 algorithm for student classroom behavior detection, named CPID-YOLOv8, is proposed. Specifically targeting the issue of small target occlusion, the Channel Prior Convolutional Attention (CPCA) was incorporated into the Backbone, enhancing the model's channel response across different scales and capturing more detailed and comprehensive contextual information. Additionally, to tackle the problem of multi-scale spanning, the Parallelized Patch-Aware Attention (PPA) multi-branch feature extraction strategy was utilized, improving the model's ability to select multi-scale features and facilitating the extraction of more effective multi-scale feature information. Furthermore, the Dimension-Aware Integration Diffusion Feature Pyramid Network (DAIDFPN) module was proposed to replace the original Neck layer for feature fusion. This module enables adaptive selection and meticulous fusion of high-dimensional and low-dimensional features. Through its diffusion mechanism, it ensures that each scale feature is provided with detailed context while reducing the model's parameter count. Experimental results demonstrate that the improved CPID-YOLOv8 model exhibits excellent detection performance in student classroom behavior tasks. Compared with the baseline YOLOv8 algorithm, the CPID-YOLOv8 model achieved improvements of 1.7%, 2.1%, 2.2%, and 2.5% in P, R, mAP50, and mAP50-95, respectively, substantiating the effectiveness of the proposed improvements.