DB-HRNet: Dual Branch High-Resolution Network for Human Pose Estimation

High-resolution networks have demonstrated significant advantages in multi-scale feature extraction for human pose estimation task. However, this often leads to the problems of large network parameter quantity and high computational complexity. On the other hand, some existing lightweight high-resolution networks have lower model parameter quantity and computational complexity, but they lack accuracy in handling keypoint position information. To address these issues, this study proposes a lightweight High-resolution network called DB-HRNet (Dual Branch High-Resolution Network) for human pose estimation. Specifically. based on HRNet, a more efficient network framework is designed as the backbone framework of DB-HRNet. Additionally, we propose a position-sensitive basic module called DBBlock (Dual Branch Block) to enhance the model’s ability to capture positional information accurately, as the basic block of DB-HRNet. Finally, we propose a Pose Refine Module to establish the connections between keypoints and make the final adjustments to the model’s output. To validate the effectiveness of DB-HRNet, we conduct experimental studies on the COCO2017 and MPII human pose estimation datasets, comparing DB-HRNet with several advanced existing lightweight human pose estimation models. The results demonstrate that the proposed network model achieves a 1.8 points improvement in detection accuracy on the COCO2017 dataset compared to the advanced lightweight high-resolution network Dite-HRNet, while also achieving a 38% increase in model inference speed.