Body Part Recognition Based On Depth Images By Learning

Depth image pattern recognition is a new technology in recent years. This is due to the cost reduction of depth image camera, especially the Kinect device introduced by Microsoft which inspires the researchers'interest greatly. Be different with optical image, pixels of depth image are only concerned with the location of objects in the space. It will not be interfered with the factors of light, shadow, chrominance and environmental change. So depth image can effectively break through the problems and bottleneck of optical image recognition. In certain space range, depth image can be used for 3D space recognition, instead of stero vision visible light cameras.This thesis mainly discusses body parts recognition by depth images. And will describe a real-time body parts recognition algorithm which is based on self-made depth image sample library, local gradient feature extraction and combined with random forest learning. Compared with the previous human recognition technology, this algorithm can detect body parts very well, and is highly accurate. Local gradient feature, extracted by this algorithm, has high discrimination to human body parts, and can overcome overlap or occlusion problems, and also is a simple calculation way. Meanwhile, as a classifier, random forest can ensure high accuracy and quick classification, and can be applied parallel computing technology which makes the speed faster. Running single thread on ordinary PC, the time of processing one image frame is 160 ms averagely, which meets the basic real-time requirement. To the training sample, the average accuracy of the classifier can reach 67.1%. And the average margin of the error between predicted and groundtruth position of body joints can be controlled within 11.5%. Through the research of depth image and the analysis of the advantages and disadvantages of the experiments, it can be known that the depth image recognition has certain development space.