基于特征融合的人体行为识别算法

Multi-features fusion algorithm for human action recognition

针对HOG特征在人体行为识别中仅仅表征人体局部梯度特征的不足,提出了一种扩展HOG(ExHOG)特征与CLBP特征相融合的人体行为识别方法。用背景差分法从视频中提取出完整的人体运动序列,并提取出扩展梯度方向直方图ExHOG及完备局部二值模式CLBP两种互补特征;利用K-L变换将这两种互补特征融合生成一个分类能力更强的行为特征;采用径向基函数神经网络RBFNN对行为特征进行识别分类。在KTH和Weizman行为公共数据库上进行了多组实验,结果表明提出的方法能够有效地识别人体运动类别。

For the inadequate of Histogram of Oriented Gradients （HOG） feature for local gradient features of the human body in human action recognition, this paper presents a recognition algorithm of human action based on multi-features fusion using extended HOG feature and Complete Local Binary Pattern（CLBP） feature. The background subtraction algorithm is used to extract the complete human motion sequence in the video, and it extracts Extended HOG and CLBP feature of human body which are complementary. Then it fuses these two group features by K-L transform to get a new feature which has a higher dis- criminating power. At last, the paper uses radial basic function neural network to realize the action of multi class classification. The experimental results in the KTH and Weizmann behavior databases show the effectiveness of the proposed algorithm.