Body Part Recognition And Action Recognition Based On Depth Image

As the depth image sensor cost reduce, the recognition based on depth image gradually becomes a hot spot in the field of pattern recognition research. Depth image is a kind of image only related to object to the sensor distance and it is not affected by external factors such as color, light and shadow. It also can effectively avoid the influence of the overlapping parts of the body in 3D space and is Very suitable for the pattern recognition in the recognition of body parts and actions.We propose a new method to quickly and accurately predict 3D positions of body joints and actions recognition from a single depth image, using no time information. Different from the traditional body part recognition and motion recognition methods, this method is per-pixel classification and don’t have to consider the connection between the front and rear frame. This paper expounds the body part recognition first. Using a large number of the depth of the synthetic image as the training data in the recognition that ensure the diversity of actions and the robustness of the result of the experiment. Firstly, we do training of random forests which consist of several decision trees through depth image feature extracting from the synthetic depth image. Secondly, we do human body parts classification based on human depth image through classifier that is a trained random forests consists of several decision trees. Lastly, we use mean-shift algorithm for clustering and skeleton joints points tracking according to the results of random forest classification, thus we do estimates of 3d coordinates about human skeleton joints. On PC, the algorithm running at a rate of about 10 frames per second can achieve real-time level. In this paper, the recognition rate of human body parts is as high as 85% of the training sample, for the untrained real depth sample recognition rate also can be 50%. On the basis of algorithm of human body part recognition, this paper extends the algorithm to the human body movements, including the recognition of single action and continuous motion. For action recognition, we also adopts the method of the classification of the random forest, the difference is class definition. On human action recognition, this paper also made a good recognition effect, recognition rate was 81%, and the course rate within 2%.This paper designes two kinds of recognition model: One is human body recognition, is the recognition of the body skeleton joints, requires a lot of training samples, and can’t determine human actions.; Another one is human action recognition, recognize daily regular actions, don’t need too much training gallery, have characteristics of real-time reality and have high application value.