In order to improve the perception ability of the walking-aid robot and make it more accurately understand the user's training status, propose an analytical method of human lower limbs dynamic model based on 3D vision for the walking-aid robot. First., we create a point cloud set of the field of view, use the advantages of 3D vision to quickly segment the leg and background point cloud data, and optimize the lower limb point cloud data through down-sampling and statistical filtering algorithms to provide accuracy and real-time guarantee for analyzing the lower limbs posture. The oriented bounding box of the local point cloud based on principal component analysis(PCA) is used to analyze the posture of the lower limbs, which is used as the angle feature of the lower limbs dynamic model. Combine the angle feature with the design constraints to obtain the dynamic model of walking lower limbs. Finally, the effectiveness of the algorithm is verified by design experiments.