Man - Made Joint Node Recognition Based On Depth Image

With Microsoft launching the Kinect depth camera, depth image pattern recognitionhas become a research hotspot. Inside, hand gesture recognition has become an importantresearch direction. As a key and basic problem, hand joint recognition of gesture recognitionis of great research significance. The information of pixel in depth image is not theinformation of brightness or color in optical image, but the distance to the camera. Therefore,compared with the optical image, depth image will not be affected by the factors such aslight, shadow, environment changes. Besides, depth image directly reflects the3Dinformation of the scene, which can provide favorable conditions for the three-dimensionaljoint recognition.This thesis puts forward a complete solution for hand joint recognition based on depthimage. Capture hand depth image by the depth camera and use the method of labeling thereal pixel to make the depth image sample database. Extract the depth features of samplesfor training random forest model and part recognition. Finally, combine information ofpixels in each part to estimate the3D position of joint point. The modified depth comparedfeatures designed in this article can effectively distinguish different hand parts and is ofrotation robustness. Random forest model has the advantages of high efficiency and parallelprocessing in the recognition problem. This system is realized on ordinary PC platform. Tothe testing sample, the average accuracy of part recognition can reach79%. And thedeviation between predicted position and ground truth position of hand joints withinallowed margin is up to82%. The time of processing one image frame by the algorithm is86ms averagely, which meets the basic real-time requirement.