Human Action Recognition Algorithm Based on Highresolution Skeletal Images

Human action recognition involves analyzing visual data of human activities to identify and understand human behaviors or postures, with wide applications in intelligent surveillance, healthcare, human-computer interaction, and other fields. Skeleton-based human action recognition algorithms estimate human poses and actions by analyzing joint points and extracting skeletal keypoint features. These methods have become a mainstream research direction in the field due to their strong anti-interference capabilities in complex environments. PoseConv3D stacks 2D skeletal keypoint maps into 3D heatmaps and employs 3D-CNNs for action recognition, achieving favorable results in both accuracy and computational efficiency. However, the PoseConv3D suffers from significant spatial information loss during skeletal keypoint extraction, leading to inaccurate joint localization. To address the information loss caused by resolution reduction in PoseConv3D, this paper proposes a human action recognition algorithm that preserves high-resolution skeletal images. The method retains high-resolution skeletal images during keypoint extraction by constructing a multi-resolution parallel subnet to preserve high-resolution feature maps during downsampling, and introducing cross-scale feature exchange units to fuse information between different subnets, thereby enhancing spatial information in skeletal keypoint images. Experiments on six datasets demonstrate that the proposed algorithm effectively improves spatial information retention for skeletal keypoints and exhibits stronger robustness compared to state-of-the-art methods.