基于图嵌入线性拓展方法的人体动作识别研究

采用图嵌入降维的方法对人侧影轮廓数据进行降维处理并用来识别人的行为动作.给定一个动作的图像序列,提取序列中每帧中人的侧影轮廓信号并用之表征人体运动,利用两种图嵌入法将提取的时变轮廓信号投影到低维空间,采用Hausdroff距离测量运动的相似性并在最近邻框架下识别人的动作.为验证算法的有效性,采用留一法和统计方法两种测试方法对五类人体常见动作(走、跑、拍手、挥手和拳击)进行测试.实验结果表明,方法不仅有很好的分类性能,而且能有效的降低了计算量.

Human Action Recognition Based on Graph Embedding

In order to recognize the human actions in the video, the high-dimension vectors were reduced into the low dimension space based on the graph embedding method. Given an image sequence of an action, the contours of the human silhouettes were abstracted and used to represent the motion of the human. Two methods of graph embedding were used to perform the low-dimensional manifold embedding of the center-boundary distance vectors of the silhouettes abstracted from the video. Five human actions (walking, running...