Methods Based On The Depth Of The Human Joints Image Positioning

Human Pose Recognition is a hotspot of current computer vision study. With the growth of computer software and hardware fields, human pose recognition technology is increasingly adhibited in game and monitoring areas. During the process of human pose analysis, poses are highly varied, and they are hidden under clothes. All these factors enhance the difficulty of human motion analysis, which brings challenges and opportunities to the growth of human pose recognition study.The birth and development of depth images brings human pose recognition technology from laboratory to people’s everyday life. Images captured by depth camera do not contain color information, but only distance information, which is not affected by factors like sunshine, shadow and etc. This, to some extend, makes it easier to recognize and forecast targets.This paper makes contributions as follows:The first chapter briefly introduces the development of human pose recognition technology and the theory of depth imaging technology, analyzes some typical human pose recognition algorithms and compares the characteristics between depth images and colorful images. The second chapter talks about the method of making segmentation between background and people using RANSAC algorithm to fit the model of ceiling and floor from collected depth images, and utilizing contour following technique and depth information of objects to extract human body areas from background images. The following chapter discusses using random forest classifier to conduct categorization for each pixel in human body area, and afterwards sort them into different body components. In particular, the method of extracting characteristics is by the help of each pixel’s offset angle information to gain local depth gradient character. This character can ensure that the pixel won’t change in the moving and rotating condition. The fourth chapter studies the way to apply mean-shift score ranking method, which aims to find the densest pixel point in probability to stand for the position of joint among the same-type body component pixels. On this basis, this paper uses priori knowledge about motion constraints to optimize the proposals of joints’ positions. The next chapter reports and analyzes results of quantitative and qualitative tests of the system and draws conclusions based on those data and compares this paper with other algorithm. The sixth chapter summarizes the system’s performance and makes a vision of this field’s futurity.The average time for the system to produce a classifier is2000s, and the precision rate is 68%, which compares to Kinect, has reached at a high level. The average time to forecast human joints" position is150ms/f. Results are basically accurate. All these prove its practicability.