基于动作捕获数据的行为识别  

Behavior recognition based on motion capturing data

作　　者：董珂[1] 甘朝晖[1] 蒋旻[2] 雷泽[1] 

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430081 [2]武汉科技大学计算机科学与技术学院,湖北武汉430081

出　　处：《计算机工程与设计》2016年第3期762-767,818,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61440016)

摘　　要：为克服人体行为识别传统方法在解决行为分割、训练集规模过小和行为无关数据干扰等方面存在的问题,提出一种基于动作捕获数据的人体行为识别方法。利用动作捕获设备Kinect获取人体骨架数据流,从骨架流数据中提取人体的姿态特征,基于决策树和动态时间规整算法进行人体行为的识别。实验结果表明,该算法可以取得较好的实验效果,比现有方法具有更好的识别准确性。To overcome the problems of behavioral segmentation,including small training set,disturbing of activity-unrelated data,an activity recognition method based on motion capturing data was proposed.Skeleton stream was obtained using the motion capture equipment Kinect,body posture features were extracted from the skeleton stream,activity recognition with decision tree and dynamic time warping was performed.Experimental results show that the algorithm can get better results with better recognition accuracy compared to the existing methods.

关 键 词：行为识别 动作捕获 KINECT 决策树 动态时间规整算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]