Classroom student posture recognition based on an improved high-resolution network

Abstract Due to the large number of students in a typical classroom and crowded seating, most features student posture are often obscured, making it difficult balance accuracy identifying postures with computational efficiency. To solve this issue, novel recognition method is proposed. First, recognize poses multiple classroom, we use you-only-look-once (YOLOv3) algorithm for object detection retrain detect human objects that hunching on table, creating pose estimation network. Next, improve network, squeeze-and-excitation network structure embedded residual high-resolution networks (HRNet). Finally, improved HRNet algorithm’s outputs key body points, design classification based support vector machine, classify classroom. Experiments show multi-person yields best mean average precision performance 73.76% common context (COCO) validation dataset. We further test proposed customer dataset collected achieved high rate 90.1% good robustness.