Instance Segmentation-Driven Pose Estimation for Automated Classroom Behavior Analysis

Accurate extraction of student posture is essential for assessing classroom dynamics and teaching effectiveness. However, in complex classroom environments, challenges such as occlusions and overlapping students make pose estimation difficult. This paper presents a novel visual extraction model that combines instance segmentation with structured pose estimation to improve skeletal extraction accuracy and robustness. Leveraging the YOLO-SEG model, we efficiently segment individual students and apply a specialized pose estimation technique using discrete token representations to capture joint dependencies accurately. This dual approach ensures precise segmentation and detailed pose reconstruction, facilitated by a robust multilevel loss function that optimizes pose accuracy and quantization consistency. Our method enhances the automated analysis of student behaviors, offering significant implications for educational assessments.