Application of Motion Image Skeleton Recognition Algorithm Based on Convolutional Neural Network in Rehabilitation Training

This paper aims to explore the application effect of Convolutional Neural Network (CNN) algorithm in motion image skeleton recognition to improve the accuracy of motion skeleton recognition and the efficiency of rehabilitation training. Here, in response to the robustness and insensitivity to changes in lighting conditions of human motion image skeleton data, the first step is to depict human motion images in real Three-Dimensional (3D) space by abstracting the human body into several joint points connected to bones to obtain 3D skeleton data. Second, the channel attention module is proposed, which is combined with the spatial graph convolution module and the time graph convolution module in the graph CNN. The 3D skeleton data samples of diverse motion images are learned to obtain good action expression ability. A motion image skeleton recognition model based on multi-attention Spatial Temporal Graph Convolutional Network is constructed. Finally, the constructed model is applied to the tracking and analysis of rehabilitation training. The results show that the model algorithm can track the patient's motor state accurately in rehabilitation training. The effectiveness of local and global information reaches more than 90%, which is obviously better than the spatiotemporal graph convolution algorithm. The model algorithm effectively reduces the problem of muscle injury caused by exercise posture error during manual intervention in traditional rehabilitation training. Therefore, the model algorithm can be widely used in rehabilitation training and practice, which provides new ideas and exploration directions for the medical industry to improve the efficiency and quality of rehabilitation training.

本研究旨在探究卷积神经网络（Convolutional Neural Network，CNN）算法在运动图像骨骼识别中的应用效果，以提升运动骨骼识别精度与康复训练效率。针对人体运动图像骨骼数据对光照条件变化具备鲁棒性且不易受其干扰的特性，本研究首先将人体抽象为若干以骨骼相连的关节点，在真实三维（Three-Dimensional，3D）空间中还原人体运动场景，以此获取三维骨骼数据。其次，本研究提出通道注意力模块（Channel Attention Module），将其与图卷积神经网络中的空间图卷积模块（Spatial Graph Convolution Module）、时间图卷积模块（Temporal Graph Convolution Module）进行融合；通过对多样化运动图像的三维骨骼数据样本进行特征学习，使模型具备优异的动作表达能力，最终构建出基于多注意力时空图卷积网络（Multi-attention Spatial Temporal Graph Convolutional Network）的运动图像骨骼识别模型。最后，将所构建的模型应用于康复训练的跟踪与分析工作中。实验结果表明，该模型算法可在康复训练过程中精准追踪患者的运动状态；其局部与全局信息有效利用率超过90%，性能显著优于传统时空图卷积算法。该模型算法有效降低了传统康复训练中人工干预阶段因运动姿势错误引发的肌肉损伤风险。因此，该模型算法可广泛应用于康复训练与实践场景，为医疗行业提升康复训练的效率与质量提供了全新的思路与探索方向。