A Category Of Human Behavior Recognition Using Skeleton And Depth Models

Recently, human behaviors recognition has being a hot topic in computer vision field. It has great influence on monitor, human and machine interaction, robot and other fields. In the several years, there comes a new equipment called Kinect, which can provide various images and audio resources. Kinect can provide color images as traditional cameras, also it can give depth image and skeleton image of humans. However, with Kinect people realized human and machine interactions and met entertainments in most situations. People overlook the ability which Kinect can bring benefits to human behaviors recognition. In the paper, Kinect is used for human behaviors recognition.In the first, the paper introduced the background, research status and common algorithms in the field of human behaviors recognition. Later the paper explained how to build a test platform for the topic, also, after necessary test the experimental environment was set. To get deep know about resources and make sure the quality of extracted features, color images, depth images and skeleton images were carefully tested in the paper. In details, the test included color image resolution, each frame rate. Also the valid of depth image and resolution were tested too. Besides, the paper considered valid range of skeleton image and sheltered situation.Features extracting is one of the most important issues of human behaviors recognition in video. Excellent features can describe human posts and behaviors well.In another word, good features are half way to success. The paper chose skeleton features, which included angles of joints, orientations, relative position, besides, features included ratios of body silhouettes. However, sheltered problems could not be avoided. Date would uncertain when came sheltered problems, to make the problem even worse, valid features cannot be extracted. So far there is no method to overcome the shortcoming. Depth images can show human posts well, and it has the advantage that date from sheltered part would not change randomly. The paper introduced HOG method for features extracting from depth images. HOG features combined with skeleton features constitute Human behaviors features.In paper, the problem of human behavior recognition was devided into two parts. One was to recognize human posts, the other was the recognition of post sequences. The well performance of SVM was proved in human post recognition. In the paper LIBSVM was introduced which was wide used. On the other hand naive bayes and HMM worked well in post sequence recognition. In the end, the paper introduced the test system based on Visual studio. The system can catch images, show images, test core algorithms, and give results. From carefully test, the method for human behavior recognition worked well. The idea in the paper had some values to application and research, also the paper provided valuable experience to later works.