基于图像识别的武术动作分解方法研究

Research on Wushu movement decomposition method based on image recognition

提出以图像识别为基础分解识别人体武术动作的方法。首先,通过形态学梯度操作使大部分噪声背景可以消除,进而取得人体轮廓边缘,将视频中每帧图像轮廓边缘提取出来并在同一幅图像中实现累积,利用累积边缘图像计算出以网格为基础的HOG,获取图像动作特征向量;其次,运用改良动态时间规整理论结合动作时间序列下各关节角度变化的特征,即可识别出各类武术动作间关节变化序列的相似性,再设计分类器并向其输入图像中人体动作时变特征数据,从而实现基于图像识别的武术动作分解过程。实验结果表明,利用图像识别可有效分解武术动作。

A method based on image recognition to identify and decompose the human-body Wushu action is proposed. The operation of morphological gradient is used to eliminate most of background noise to obtain the edge of the human-body contour. The contour edge of each frame image is extracted in the image, and accumulated in the same image. The cumulative edge image is used to calculate the HOG based on grid, and acquire the action feature vector of the image. The improved dynamic characteristics of time warping theory is combined with the change characteristic of the angle of each joint under movement time series to recognize the similarity of joint change sequence among various Wushu movements. The classifier was designed, and the time-varying characteristic data of human-body movement in the image is input into it to realize the Wushu movement decomposition based on image recognition. The experimental resuhs show that the image recognition can decompose the Wushu movement effectively.