A High Efficient Training Method for Student Action Detection from Few Pixel in Complex Crowded Scenes

Student action recognition plays a crucial role in assisting teachers to provide targeted support to students. Common classroom action detection systems struggle with accurately tracking and detecting multiple targets in dense student crowds, and the accuracy of learning behavior recognition is insufficient to meet the demands of precise classroom behavior identification. To address these issues, a highly efficient training method for student action detection from a few pixels in complex crowed scenes is proposed in this paper. The proposed framework consists of four modules: Dense Scene Face Localization, Fixed-point Student Segmentation, Efficient Labeling, and Classroom Behavior Detection. The Dense Scene Face Localization module identifies the facial regions of all individuals in the classroom. The Fixed-point Student Segmentation module segments the pixels occupied by each student to support efficient labeling. The Efficient Labeling module generates the necessary labels for the Classroom Behavior Detection module, which then detects the behaviors of individuals in the classroom. Our data is sourced from surveillance videos of teaching classrooms at Chongqing University of Posts and Telecommunications. Our proposed framework substantially reduces labeling costs. The experiments show that for the student looking up/down 2 category action recognition task in complex scenarios, only 293 (29.3%), 382 (19.1%), 665 (16.6%), 1116 (14.0%) and 4857 (24.3%) manual labeled images are required to train on 1000, 2000, 4000, 8000 and 20000 images, respectively and finally obtained average prediction accuracy performance of 92.15% on 3142 test image datasets, and the prediction accuracy on test datasets of varying image size with super small (<5KB), small (5KB~7KB), median (7KB~9KB), and large (>13KB) image size is 92.2%, 93.3%, 93.7% and 89.4% respectively. This method offers a high efficient approach to recognizing student behaviors in complex classroom scenarios characterized by multiple individuals, occlusions, and small target recognition. Systematic experiments have validated the algorithm's feasibility.