结合区域HOF和字典学习的人体行为识别方法

Human activity recognition method combined with region HOF and dictionary learning

为辅助老人看护,提出一种基于视频的人体行为识别方法。采用AMBER方法检测视频中的运动目标,粗定位行为感兴趣区域;提取行为感兴趣区域各像素点的光流,构建归一化的光流方向直方图(histograms of oriented optical flow,HOF),用于描述人体行为;采用在线字典学习方法进行训练和测试,在训练阶段寻找最优的字典和稀疏矩阵,在测试阶段依据稀疏性分类不同特征。在国际上通用的ADL人体行为数据库中的仿真实验结果表明,采用本方法进行人体行为识别的识别率高,且不同人体行为之间的分类混淆率低。

As a secondary care for the elderly, this paper proposed a new human activity recognition method based on video. First, it used AMBER method for detecting the moving objects in video, and located the activity region of interest roughly. Then, it extracted optical flow of each pixel in the activity region of interest, and built normalized histograms of oriented opti- cal flow （HOF） , which was used to describe human activities. Finally,it used online dictionary learning method for training and t^sting, to find the best dictionaries and sparse matrices during training phase, and classified different features according to sparsity during testing phase. The results of experiment on the international human activity dataset show that, recognize human activity by using the new method can achieve high recognition rate, and low category confusion rate between different human activities.：