In order to achieve the evaluation of human rehabilitation training movements, a human 3D pose estimation network integrating key-frame enhancement method (KFEM) and CTRAMM module is proposed, and a matching algorithm based on location and type dynamic time warping (LTDTW) is developed to evaluate rehabilitation movements. KFEM determines key-frames and adjusts their weights by calculating the coordinate transformation of human key-points. The CTRAMM module dynamically learns different topological structures, improving the feature representation ability of the model. The LTDTW improves the accuracy of sequence matching through adaptive weight coefficients. The experimental results on different datasets have validated the effectiveness of the proposed method.