摘要
在词袋模型基础上,综合考虑姿态之间的时序约束关系,提出一种基于局部匹配窗口的动作识别方法.首先采用人体姿态差别作为动作序列特征描述.其次,在模型学习阶段,使用局部训练法而非传统的整体训练法来提高特征词汇的表征性;在特征量化阶段,使用自适应局部线性重构策略来给特征基更灵活的权值;在对象描述阶段,分别使用时间金字塔、滑动窗口2种方法将整个动作序列划分成多个局部动作片段,进而通过连接各个局部动作片段的特征来描述整个动作序列.最后使用直方图相交操作来完成特征匹配工作.在MSR Action3D数据库上测试了所提算法的性能并与已有的动作识别方法进行对比,结果表明,该方法的识别效果较优.
Based on the traditional bag-of-words model, we propose a new human action recognition method based on local window matching strategy which considering the temporal and spatial constraints among human pose. Firstly, human posture difference is used as features of human action sequences. Secondly, in the training phrase, instead of the global training method the local training method is used to improve the characterization of traditional characteristic vocabulary. In the feature quantization phrase, we use the adaptive local linear reconstruction strategy to give feature vectors more flexible weight. In the object description phrase, we use the time pyramid method and the sliding window method respectively to segment the whole action sequence into a set of local motion fragments, and then features of these local motion fragments are connected to describe the whole action sequence. Finally, we use the histogram intersection operation to do feature matching work. Experiments are done on the MSR Action3D database. The results show that our method can get better recognition ratio.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2014年第10期1764-1773,共10页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(61303142
61173096
61103140
61201074)
浙江省自然科学基金(Y1110882
Y1110688
R1110679)
教育部博士点基金(20113317110001)
浙江省教育厅一般科研项目(Y201330304)
关键词
动作识别
局部匹配窗口
词袋模型
深度图像
human action recognition
local window matching
bag-of-words
depth image
