Acceleration And Performance Analysis Of Binocular Stereo Matching Algorithm With Feature Fusion

Binocular stereo vision is an important means for machines to understand the objective world.Binocular stereo matching is an important part of binocular stereo vision technology,which aims to extract 3D information of the scene from 2D image data to achieve a comprehensive understanding of the 3D world.With the advantages of low cost,easy operation and high reliability,binocular stereo matching is widely used in robot navigation,virtual reality/augmented reality,autonomous driving obstacle avoidance,and non-contact distance measurement.The above application scenarios need to obtain accurate 3D information of objects in a very short period of time,therefore,the real-time and accuracy of the matching algorithm is more demanding.In this paper,we focus on both algorithm design and hardware platform to improve the operation efficiency of binocular stereo matching algorithm.The algorithm design level is to study and design a stereo matching algorithm with low computational complexity and high matching efficiency;the hardware platform aspect is to utilize the parallel computing capability of GPU platform to realize real-time processing for the binocular stereo matching algorithm.Finally,the algorithm obtains obvious acceleration effect and effectively solves the problems of real-time processing and accuracy of binocular stereo matching.The main research works of this paper are as follows.1.To address the defects that the traditional Census transform algorithm relies too much on the central pixel value and is sensitive to image noise.The sum of Absolute Differences(SAD)of pixel grayscale differences is used as the initial matching cost with the result of feature fusion of Census transform algorithm.Then,the multi-path SGM algorithm is used for cost aggregation,the optimal parallax is calculated by using the classical "winner-take-all" strategy,and finally the occlusion points are detected and filled by the left-right consistency check,and the outliers are removed by the median filter in the above process,and the optimized parallax map is finally obtained.The experimental results show that the fusion of the two algorithms can not only improve the matching accuracy,but also maintain good stability for the image brightness differences caused by lighting changes,facing repetitive local regions and large texture-free regions.The subsequent series of image optimization processing algorithms effectively improve the accuracy of the matching algorithm in this paper.2.To address the problem of time-consuming binocular stereo matching algorithm,this paper adopts Compute Unified Device Architecture(CUDA)to achieve parallel and accelerated computation for the algorithm.The algorithm takes full advantage of GPU multithreading and multi-core parallel computing,maximizes the use of shared memory and register memory,and uses CUDA stream operations to parallelize between different core functions,which greatly improves the matching efficiency of the algorithm.As shown by the experimental results,the algorithm reduces the full-region average mismatching rate by 8.05%on the Middlebury stereo matching platform;runs 450×375 resolution images on the NVIDIA Ge Force GTX 1650 platform,which is 687 times faster than the classical SGM algorithm;runs high-resolution images(1241×376)when the parallax range is 128 When the parallax range is 128,the real-time display performance(112.4fps)can still be achieved.At the same time,the accuracy and real-time performance of our algorithm are also outstanding when facing complex outdoor road scenes and indoor scenes under natural light.Our research results have wide application prospects and practical applications in the field of real-time binocular vision.