In recent years, human action recognition has a broad application prospect in intelligent family, intelligent monitoring, virtual reality and other fields. Especially in the intelligent family system, people want to control home appliances in a more intelligent way. The precondition to realize intelligent control is that computer can capture the human posture parameters and identify movements to understand people’s intentions, if people want to capture people’s posture parameters, that is to say, people need to obtain the location of the key points. Initially, the research on the positioning of key points is based on visible light imag e to expand, but visible light image is susceptible to environmental change such as color, light and shadow, it make the study of the key points subject to many limitations. Microsoft campany launched Kinect sensor in 2010, it greatly stimulates the interest of the researchers. Unlike visible light image, the gray value of every pixel in the depth image is only related with the distance between the objects and Kinect, the cost is lower than other depth sensors, it also can drastically solve the problems and bottlenecks of visible light image research. At the same time, the depth image can estimate the human body posture of three dimension in a certain range.Therefore, this paper uses Kinect to obtain the human depth images, study and implement the human body joint point localization algorithm, which is suitable for the intelligent family system.The main research content is as follows:This article uses the depth sensor to obtain the human depth images and get the depth value of each pixel. It is the basis of further research; Image pixel classification is the focus of this article. This article divides the human body into 26 different parts, we use different color to mark the corresponding parts and extract local differential feature of every part, to identify each pixel’s class by the learning of random forest; The positioning of joint point is the ultimate goal of this article. We find the center of mass of every part to position 15 joint points by using the average mass center theory.Different from the previous positioning technology of the human body joint point, the algorithm can accurately detect the various parts of the body and find the location of the joint points. The differential feature of the algorithm is simple to calculate, it can effectively distinguish between different parts of the body, to a certain extent, it also can solve the problems of integration and cover of body. Random forest classifier can quickly and accurately realize the classification of image pixel point by point. |