Research On Technology Of Moving Human Detection Based On Kinect

Abstract:Human detection plays an important role in the area of public security and building of smart city, and it is the premise of recognition of human identification and action. Meantime, the widespread use of Kinect which is typical of stereo vision, make the detection of moving human bodies from videos conveniently and exactly. Therefore, the research of detection of moving human bodies based on Kinect is significant to practical application.In the application of detection based on Kinect, following work has been finished in this research:(1) Because Kinect is a binocular vision system, before using it to collect videos, the calibration of cameras should be done firstly. Then compute the rotational matrix and translational vector between two cameras. Additionally, a conclusion that the discrepancy between FOV(field of vision) of IR and RGB camera has greater influence than AOV(angle of view), is obtained, thus, the coordinate system transformations between range images and RGB images cannot overcome the mismatching problem.(2) Research of detection algorithms. Firstly, introducing the single Gaussian model, an automatic updating Gaussian model(AGM) is used to construct background one. Thresholds are calculated by the statistical values of AGM, then foreground targets are caught by thresholds. Secondly, because of jump between the area of a body and background in range images, sobel computer is utilized to detect human bodies. Thirdly, using the conclusion of calibration and imaging, uniting the RGB images and depth one, a zooming registration detection algorithm is proposed. Lastly, evaluation of these algorithms, such as time consuming, integrity of detected body area and adaptivity of algorithms, is implemented in quantitavely.(3) Researches after detecting the foreground. Far distance and res pectivity of near infrared to different materials cause great loss of depth data and noise. In this paper, inpainting rules and models based on pa rtial differential equation are introduced firstly, and acc-ording to the ra nge image inpainting problem, neighborhood vertical gradient inpainting method is proposed. This algorithm uses the st-atistical information of local area near lose one as the reference and PDE model as rectificatio n model, to realize the inpainting. Ex-periments show that it can achiev e an exact result. Laterly, unifying structural information of2D and3D depth value, by computing mean depth value of respective local area i n2D depth image, a filter is got to remove noise effectively.(4) On the platform of Microsoft visual studio2010, a prototype system which can collect videos of moving humans and detect them from videos, is constructed by MFC modules.