Research On Realization Of Stereo Matching Algorithm Based On Embedded GPU

Stereo matching is a technology which obtain parallax image by using binocular camera.It`s a hot spot in the field of computer vision research and play an important role in robot navigation,automatic driving,the scene of three-dimensional reconstruction and other aspects.The large computational complexity of the stereo matching algorithm makes its processing speed very slow even on the PC,the realization of the PC is difficult to mean the actual requirements of power consumption,size and price.Based on Nvidia's embedded GPU development platform Jetson TX1,we use CUDA technology to realize the parallel design of the matching method based on guided filtering,and propose an improvement to the semi-global stereo matching algorithm.In this paper,the CUDA platform is introduced on the basis of the research history and current situation of stereo matching algorithm.The binocular camera model of the binocular camera is calibrated and corrected by Zhang Zhengyou calibration algorithm.The input image of the stereo matching is subjected to the line correction and the mean filter preprocessing.Then we use the adaptive support weight algorithm as a contrast to study the guided filtering method in detail.Combined with CUDA,we complete the parallelization design of cost calculation and cost aggregation phase of the boot filter matching method,and realize the processing speed increase of 37.8 times.Finally,we improve the semi-global stereo matching algorithm,and propose an AD-Census matching cost function which is improved by using the cross region.Based on the multi-scale information,the cost aggregation is carried out according to the multi-scale model.Our experiments show that the improved semi-global matching method achieves a parallax image with a clearer edge while the average bad point rate is also reduced by 2.2%.