Human Body Pose Recognition From The Depth Image

Human pose recognition is an important research direction in computer visionand it has broad application prospects in many fields, such as intelligent monitoring,advanced human-computer interaction, human motion analysis. However, becauseordinary optical images are easily influenced by outside variables such as light,shadow, the study of the human pose recognition has not been made the breakthrough.Until recent years, with the development of the depth sensing device, especiallyMicrosoft launched Kinect depth sensing device, human body pose recognition fromthe depth image has greatly stimulated the interest of researchers in the field ofcomputer vision. So far, there are already many, of which some algorithms haveobtained better results. Among advanced algorithms, it is more influential that Shottonet al have achieved real-time human motion capture with Kinect. The capture systemmeets the real-time requirements that human has been requiring to some extent, butthe algorithm is not suitable for low-resolution depth image of the human body poserecognition, at the same time, the algorithm needs the higher hardware requirementsof the training platform too. So，in order to achieve a higher accuracy of humangesture recognition from the rate low resolution depth image，the paper builds thesmall data including the depth image and gray image of body posture，using computergraphics technology and existed database. And the paper detects human body partsusing random forests and predicts the joints of every parts on a regular PC.The main research work has the following two parts:1) Given there is no public depth database of body posture, the paper builds thedepth of the data sets containing low-resolution images of the human bodyposture， with the use of Maya platform based on the computer graphicstechnology and existing CMU motion capture database. Compared to the depthsensor for depth image manually and then individually marked parts of the humanbody, this method not only saves manpower and effort, but human error is smalland difficult to produce classified disaster, more importantly, is covering thecommon people’s daily movement gesture sequence.2) The thesis detects human body parts using a classic random forest classificationmodel, and we use the depth differential characteristic of pixel of depth image asthe characteristics to train the random forest. We use the he weighted Mean Shiftalgorithm to search mode of every of human body parts after some parts beingmerged, and the mode is joint point of human body part.Theory and ideas of data of the body pose depth image to later computervision tasks of the depth image-based research provides some reference value.