A Lightweight and Efficient Algorithm for Human Pose Estimation

Human pose estimation is a critical task in the field of computer vision, aiming to identify and localize keypoints of the human body by integrating visual image information with deep learning techniques, thereby obtaining global pose information of limbs and torso. However, current mainstream human pose estimation algorithms often suffer from issues such as a large number of parameters, high computational requirements and insufficient feature extraction capabilities. To address these challenges, this paper proposes FCTransPose (Transformer-based Pose Estimation with Feature-fusion Pyramid Block and Criss-Cross Attention), a lightweight and efficient human pose estimation model based on TransPose (Transformer-based Pose Estimation). This model introduces the CCA (Criss-Cross Attention) in the Transformer encoder layer and improves the feedforward network layer, effectively reducing computational complexity and improving network efficiency. Additionally, an efficient multi-branch feature fusion module, FPB (Feature Pyramid Branch), is proposed and integrated into HRNet (High-Resolution Network), further enhancing the model's localization accuracy and reducing the parameter count. Experimental results demonstrate that the proposed model achieves an average precision of 76.2% and an average recall of 78.9% on the COCO (Common Objects in Context) validation set, representing improvements of 2.0% and 0.9%, respectively, compared to TransPose. Moreover, the parameter count is reduced by 40%, and the computational cost is decreased by 60%, validating the effectiveness and superiority of the proposed method. In the future, this model is expected to enable efficient human pose estimation in resource-constrained scenarios.