基于插值小波关键帧提取的动作评价算法

Action Evaluation Algorithm Based on Key Frames Extracted by Interpolating Wavelet

人体运动的行为特征具有多样性和复杂性,在运动的不同阶段有些动作的剧烈程度差异较大,但现有方法在进行动作相似度评价时未充分考虑该因素,使得评价结果存在一定偏差。针对该问题,基于多尺度FaberSchauder插值小波对参考动作序列中运动最剧烈关节的四元数分量时间序列分别提取关键帧。通过合并4组关键帧,设置阈值剔除相似度较高的关键帧。采用动态时间规整方法对参考动作和对比动作进行匹配,得到对比动作序列的关键帧,将2组关键帧的平均距离归一化后作为动作相似度评分。实验结果表明,提出的算法能够较好地实现动作评价,且对于较相似的动作,也能获得较好的评价结果。

Characteristics of human motion are always diverse and complex.The intensity of some actions is quite different at different stages of the movement,but the existing methods don＆#39;t take this factor into account when evaluating the similarity of actions,which makes the evaluation results not accurate enough.To address this problem,this paper extracts key frames of four component time series of the most violent joint from reference action sequence by multi-scale Faber-Schauder wavelet interpolation.The four groups of key frames are merged and a threshold is set to exclude the key frames of high similarity.Dynamic Time Warping （DTW） method is used to match the reference action and contrast action to get key frames of the contrast action.The motion similarity score can be calculated by normalizing the average distance between key frames of the two actions.Experimental results show that the proposed method can achieve a better evaluation than other methods and the evaluation result is also better for some similar actions.