基于图像的动作识别与评价方法

An Image-Based Action Recognition And Evaluation Method

为了为体育运动提供科学的辅助训练与实现公正的打分,利用图像识别技术,提出一种基于图像的动作识别与评价方法。该方法首先对动作视频进行图像帧提取,通过滤波消除图像噪声;然后通过OpenPose姿态估计算法提取人体骨骼关节点坐标,根据关节点坐标对人体骨架运动特征进行特征向量的构建;最后利用动态时间规整算法和KNN分类器完成动作识别和评价。经实验验证,该方法在自制数据集的8类太极拳动作以及KTH和Weizmann两个公开数据集中表现出良好的效果。具有一定实用价值。

In order to provide scientific auxiliary training for sports and realize fair scoring, an image-based action recognition and evaluation method is proposed by using image recognition technology.Firstly, the image frame of the action video is extracted, and the image noise is eliminated by filtering.Then, the coordinates of human skeleton joints are extracted by OpenPose attitude estimation algorithm,and the feature vectors of human skeleton motion characteristics are constructed according to the coordinates of joints. Finally, the dynamic time warping algorithm and KNN classifier are used to complete the action recognition and evaluation. The experimental results show that the method has a good effect on eight kinds of Tai chi actions in self-made data sets and two public data sets, KTH and Weizmann. It has certain practical value.