Research On Parallel Stereo Matching Based On GPGPU And Application System

Stereo matching is the most important part of stereo vision system.It has wide applications in society,such as 3d reconstruction,driver assis-tance,aviation area and so on.So it's meaningful to do research on stereo matching.Now stereo matching algorithms has two problems.The first one is the algorithm performs badly when the scene is disordered.The second one is the algorithm needs much computation,thus wasting much time.Our research mainly focuses on these two points.The main jobs in-clude:Firstly,analyze and improve traditional stereo matching algorithms.We study by four steps:matching cost computation,cost aggregation,disparity selection and post-processing.Then we point out the shortages and give an improved solution.In matching cost computation step,we present a novel method based on sobel operator and census transfor-mation.In cost aggregation step and post-processing step,we use rolling guidance filter to eliminate the influence of small scale structures when computing weights.We also present a novel method when the original stereo images' resolution is large,thus saving much time but adding a lit-tle more errors.Finally we use Middlebury benchmark to prove our im-provement.Secondly,we present that we should use GPGPU to accelerate stereo matching algorithm.We point out that integral image algorithm is the key step and we investigate traditional implementation.We propose a novel implementation based on CUDA texture memory and CUDA shuffle in-structions.Thus,we can reach 10 FPS or above to fulfill the real-time re-quirement.Besides,we propose a novel solution based on two same GPUs.It can save about one third time compared with one GPU solution.We utilize NVIDIA Tesla K20c cards and CUDA 7.0 platform to prove our improvement.Finally,we design and implement a 3D information processing sys-tem based on GPU and stereo matching algorithm.The system can meas-ure distances,assist driving,do living face detection and process digital images.The system has good speed for GPU accelerating.The experi-ments prove the effectiveness of our system.