基于动态视觉信息归一化的健身动作分析系统

Fitness action analysis system based on dynamic visual information normalization

基于视频序列的动作识别和评估技术已广泛应用于智能健身和主动健康领域,然而由于健身场景的复杂性,拍摄采集的视频画面通常不如实验室条件下采集的稳定,相机移动带来的视角变化使得动作识别和评估的精度受到影响。为满足智能化健身的应用需求,提出一种基于动态视觉信息归一化的健身动作分析系统,以便捷实现自由移动机位下的健身动作分析。首先,使用基于深度学习框架的人物分割模型、相机位姿估计方法提取健身视频中的静态图像特征估计相机位姿;再结合相机位姿信息,基于几何投影关系建立单目相机-人体运动关系模型;最后,将人体姿态变换到同一坐标系下,基于归一化人体骨架信息建立轻量级机器学习模型进行动作识别与动作质量评估。实验结果表明,相机-人体运动关系模型和人体骨架归一化方法能有效处理移动机位拍摄的动作,解决人体健身动作自遮挡等问题并完成动作识别。基于骨架编码信息的余弦相似度方法能有效完成健身动作评价工作。在完成以上两项工作的过程中,Kinect提取的人体骨架质量明显高于OpenPose,动作识别精度较OpenPose骨架提高12.5个百分点。所提系统能在移动视角下有效识别健身动作并进行质量评估,为改善动作表现提供量化依据。

The technology of action recognition and evaluation based on video sequences has been widely used in intelligent fitness and active health fields.However,due to the complexity of fitness scenes,the captured videos are often less stable than those captured under laboratory conditions.To solve the problem and meet the application needs of intelligent fitness,a fitness action analysis system based on dynamic visual information normalization was proposed,which can easily realize the analysis of fitness actionin free-moving camera positions.Firstly,a depth-based learning framework was used to extract static image features from fitness videos for camera pose estimation.Then,combined with the camera pose information,a monocular camera-human motion relationship model was established based on geometric projection.Finally,the human pose was transformed to the same coordinate system,and a lightweight machine-learning model was established based on normalized human skeleton information for action recognition and quality evaluation.Experimental results show that the camera-human motion relationship model and the normalization method of the human skeleton can effectively deal with the movement of the camera shooting and solve the problems of self-occlusion and other issues in human fitness action to complete the action recognition.The cosine similarity method based on skeleton coding information can effectively complete the evaluation of fitness action.In the process of completing the above two tasks,the quality of the human skeleton extracted by Kinect is significantly higher than that of OpenPose,and its action recognition accuracy is 12.5 percentage points higher than that of OpenPose.The proposed system can effectively recognize fitness action and evaluate their quality in mobile perspectives,providing a quantitative basis for improving action performance.