中文摘要：

针对一般聚类获得的码本缺乏判别性表示导致不能有效进行人体动作识别的问题,提出了一种新的自适应码本学习方法,该方法将判别式词袋（bag of words,Bo W）动作表示和自适应码本学习结合,增强了码本的表示能力和特征的判别性。为了有效求解非凸目标函数,提出基于轮换优化迭代方法,即固定码本更新判别矩阵,然后判别矩阵更新固定码本,直至满足终止迭代条件,该方法为自适应码本学习提供了技术支持。仿真实验采用KTH、Hollywood2、芭蕾、i3Dpost数据库进行判别比较,识别率比现有典型方法平均提高了4%左右,学习到的码本在特征空间中具有良好的判别性能。相比于基于光流、方向梯度直方图（histograms of oriented gradients,HOG）等方法,计算复杂度更低,实用性更好。

英文摘要：

As the problem of ineffective human action recognition caused by lack of judgment representation of codebook obtained by general clustering,this paper proposed a new adaptive codebook learning method,combined discriminant bag of words with adaptive codebook learning,which had enhanced the ability to represent features of codebook discriminant. In order to effectively solve the non-convex objective function,it proposed iteration of rotational optimization,that was the fixed codebook updated the judgment matrix,and then the judgment matrix updated the fixed codebook until it met the conditions for termination of the iteration. This method provided technical support for the adaptive codebook learning. In the simulation experiments,it tested five data sets KTH,Hollywood2,ballet,i3 Dpost and facial expressions. The experimental result is about 4% recognition rate averagely higher than the existing typical methods. The learned codebook has good discrimination performance in the feature space. Compared with optical flow method,the method based on the histograms of oriented gradients（ HOG） and other methods,the proposed method has lower computational complexity and better usability.