3D pose recognition method for rehabilitation training based on the fusion of CBAS-HRNET and stereo vision

To address issues such as low accuracy and inconvenient wearability in traditional rehabilitation training systems for patient posture recognition, this paper proposes a three-dimensional pose recognition method that integrates high-resolution networks and stereo vision, aiming to enhance both the accuracy and efficiency of movement ability recognition. An improved novel network architecture based on the high-resolution network (HRNet) is introduced, which incorporates a mixed attention module at the feature extraction layer to reduce detail feature loss from convolutional block attention module (CBAM). Additionally, the atrous spatial pyramid pooling (ASPP) module is employed at the feature fusion layer to strengthen the model’s ability to extract multi-scale feature information. The predicted 2D poses from the left and right images are used as matching feature points, and stereo vision systems calculate the 3D coordinates to obtain the posture data for rehabilitation training movements. Experimental results demonstrate that the improved 2D pose estimation model achieves an average accuracy of 76.1% on the COCO dataset, a 1.7% improvement over the original model. In the 3D pose recognition experiment for spinal rotation capability, the measured average relative error of rotation angles is 4.18%, meeting the expected detection performance.