Study Of The Stereo Matching Algorithm Based On Adaptive Census And FPGA Implementation

With the development of artificial intelligence,the research of computer stereo vision system has attracted more and more attention of many researchers. A common stereo vision system includes camera calibration,images acquision,stereo matching and 3D reconstruction.Stereo matching plays the most important role in the stereo vision system.The stereo matching algorithm with high matching precision and low complexity always is the study hotspot. Based on the stereo matching constraint and disparity principle, Stereo matching gets the depth information of the target using the images of the same target which is taken by the binocular camera from two different perspectives.The development of stereo matching algorithm is introduced and the advantages and disadvantages of some kind of algorithm are analyzed in this article.Based on the traditional Ceneus transform algorithm, aiming at the limitation of low accuracy in the region where the depth is not continuous of the the traditional algorithm, a novel stereo matching algorithm for adaptive weight Census transform is proposed. A absolute value derived from the central pixel and its cross-mean is to determine the central gray value in the stage of Census transform;With the purpose of making a difference with each pixel in the transform window,a method using linear segmentation to calculate the adaptive weight is used. The window size in the stage of cost aggregation is determined by comparing the gradient value of the center and its the two nearest points. The experimental results show that the matching accuracy of the proposed algorithm is much better than the traditional Census algorithm, especially in the discontinuous depth region, the complexity of calculation and the difficulty of hardware implementation is not increased obviously.In the aspect of hardware implement,the original stereo matching system based on Digilent Atlys board and VmodCam Module of binocular camera is designed in the paper.And the system has achieved the modules,such as VmodCam control module, image storage management module,transmission module and clock management module.Besides,the proposed stereo matching algorithm has been implemented initially.The paper gives the realization of the weights calculation,Census transforma-tion, matching cost calculation and aggragation and shows the preliminary design of parallel structure.By the test of the initial system in the real scene,it is proved that the matching accuracy of the proposed algorithm is much better than the traditional Census algorithm, especially in the discontinuous depth region.