Adaptive Target Region Attention Network-based Human Pose Estimation in Smart Classroom

Abstract

In smart classroom environments, problems such as occlusion and overlap make the acquisition of student pose information challenging. To address these problems, a lightweight human pose estimation model with Adaptive Target Region Attention based on Lite-HRNet is proposed for smart classroom scenarios. Firstly, the Deformable Convolutional Encoding Network (DCEN) module is designed to reconstruct the encoding of features through an encoder and then a multi-layer deformable convolutional module is used to adaptively focus on the image region to obtain a feature representation that focuses on the target region of interest of the student subject. Secondly, the Channel And Spatial Attention (CASA) module is designed to attenuate or enhance the feature attention in different regions of the feature map to obtain a more accurate representation of the target feature. Finally, extensive experiments were conducted on the COCO dataset and the smart classroom dataset (SC-Data) to compare the proposed model with the current main popular human pose estimation framework. The experimental results show that the performance of the model reaches 67.5(mAP) in the COCO dataset, which is an improvement of 2.7(mAP) compared to the Lite-HRNet model, and 86.6(mAP) in the SC-Data dataset, which is an improvement of 1.6(mAP) compared to the Lite-HRNet model.