Research Of3D Reconstruction And Filtering Algorithm Based On Depth Information Of Kinect

3D reconstruction is one of the research hotspots and difficulties in theareas of computer vision, artificial intelligence, virtual reality and so on. Toimplement3D real-time network conference system, this article proposed3D reconstruction algorithm using the depth information of Kinect, andreconstructed3D model in real-time.First of all, this article introduced structure of Kinect, and analyzed theprinciple of achieving depth information. At the same time, it alsointroduced the principle of camera calibration, geometry relationship inimaging, the model of camera distortion and geometry relationship of twocameras. After introduction the principle, this article designed the calibrationprogram for Kinect, and applied it to both the color camera and infraredcamera. Due to the laser point interference to infrared image, preliminarycalibration results for infrared camera had a big error. According to that, thisarticle proposed two approaches, shielding infrared transmitter and settingthe calibration environment to sunny room. After that, the error ofcalibration was reasonable. The final results show that radial distortions ofboth the color camera and infrared camera have been calibrated insideKinect, and tangential distortion is too small to calibrate. And the twocameras only have translation in one direction, and the rotation is too small.That can be proved from the appearance of Kinect.After comparing four typical3D reconstruction algorithms, this articlethen described real-time3D reconstruction algorithm based on the depthinformation of Kinect. According to the noise problem, this article analyzedthree sources of noise, modeled random noise, and did experiment toapprove that. Based on the signal structure of depth image and noise, thisarticle applied existing median filter, Gaussian filter, bilateral filter, jointbilateral filter, and self-improved Gaussian filter and joint bilateral filter, toprocess the original depth image, and analyzed the filter effect according to 3D cloud points and depth image. To apply bilateral filter, this articlestudied how to select right parameters to achieve the best de-noising effect,using error of mean square as criteria. At last, this article analyzed thealgorithm performance, and proposed high-performance algorithm.Compared to previous reconstruction algorithm, the algorithmproposed in this article can reconstruct more-refined3D models, and almostachieve real-time performance. But it isn’t fast enough, still needs furtherboosting using GPU boosting technology and so on.