Human Pose Estimation Algorithm Based on Improved Lightweight High-Resolution Network

Human pose estimation is of significant importance in fields such as autonomous driving, rehabilitation therapy, and motion capture. Addressing the computational load challenges faced by current pose estimation models when improving detection accuracy, this paper proposes a lightweight pose estimation model called Lightweight Spatial Aware High-Resolution Network (Lite SA-HRNet), based on the Lightweight High-Resolution Network (Lite-HRNet). Initially, the Conditional Channel Weighting Module is optimized by incorporating the Positional Convolutional Block Attention Module (PCBAM) in cross-resolution weight calculations, enhancing the feature map’s expressiveness and the model’s accuracy. Additionally, the Spatial Context Aware Module (SCAM) is introduced at the final stage of feature fusion to further differentiate between targets and irrelevant backgrounds. Experimental results on the COCO2017 dataset demonstrate that Lite SA-HRNet has 1.93M parameters and 0.28 GFLOPs of computational complexity, with a $0.9 \%$ improvement in average precision (AP). This validates that the proposed network effectively maintains a balance between lightweight design and detection accuracy.