结合HOG/HOF级联特征和多层分类器的人体行为识别

Human activity recognition combined with HOG/HOF features and multilayer classifier

提出一种人体行为识别方法,结合HOG/HOF级联特征和多层分类器提高人体行为的识别率。采用VIBE方法提取视频中的前景区域;在前景区域上分别提取方向梯度直方图(HOG)特征和光流方向直方图(HOF)特征,组成HOG/HOF级联特征,将视频片段中所有前景区域的HOG/HOF级联特征构建成一个特征向量集合;构建包含两层自组织映射网络和一层有监督神经网络的多层分类器,对视频片段的特征向量集合进行分类,得到行为识别结果。仿真结果表明,该方法的行为识别率高,对不同人体行为的分类混淆率低。

A human activity recognition method was proposed to improve recognition rate for human activities by combining HOG/HOF features and multilayer classifier. VIBE method was used to extract the foreground regions in the video. Both his-togram of oriented gradients （HOG） and histogram of oriented optical flow （HOF） features of these foreground regions were ex-tracted respectively, and a set of feature vectors with HOG/HOF features of all foreground regions in a video clip was construc-ted. A multilayer classifier containing two layers of self-organizing maps network and one layer of supervised neural network was built to classify the set of feature vectors for obtaining a result of activity recognition. Simulation results show that the proposed method has high recognition rate, and low category confusion rate for different human activities.