基于改进OpenPose的视频动作评分方法研究

Research on Video Action Scoring Method based on Improved OpenPose

针对人体姿态估计模型OpenPose计算量大、检测速度慢等问题,提出了一种改进OpenPose模型,替换其主干网络为八度卷积与MobileNet融合而成的Oct-MobileNet,并优化缩减预测阶段的重复分支。实验表明,改进模型的计算量降低为原来的12%且检测速度提升300%。应用改进OpenPose模型提取标准视频与测试视频的姿态向量时间序列,其中姿态向量由关键点坐标经归一化处理后组合得到。采用姿态向量之间的余弦距离表征单帧动作相似度,通过动态时间规整算法计算标准序列与测试序列之间的累积距离作为序列整体相似度。该评分方法计算复杂度低且适用于视频时长不一致的情况,在八段锦健身动作评估中取得了较好应用效果,具有一定的推广应用价值。

Aiming at the problems of large amount of calculation and slow detection speed of the human body posture estimation model OpenPose,this paper proposes an improved OpenPose model,the backbone network of which is replaced with Oct-MobileNet,a fusion of octave convolution and MobileNet.The duplicative branches in the prediction stage are optimized and reduced.Experimental results show that the calculation amount of the improved model is reduced to 12%and the detection speed increases by 300%.The improved OpenPose is used to extract the pose vector time series of standard video and test video,in which the pose vector is obtained by combining the key point coordinates after normalization.The cosine distance between the pose vectors is used to characterize the action similarity of a single frame,and the cumulative distance between the standard sequence and the test sequence is calculated by the dynamic time warping algorithm as the overall similarity of the sequence.The proposed scoring method has low computational complexity and is suitable for the case of inconsistent video duration.It has achieved good application results in the evaluation of Baduanjin fitness movements,and has certain promotion and application value.