Research On Stereo Matching Algorithm Based On GPU

Stereo vision based driver assistance system (DAS) is a significant task in the field of computer vision. The technique simulates human visual mechanism, percepts the three-dimensional information through stereo images captured by two cameras, assists the driver through road information computed by computer in the vehicle. Data processing in DAS must be synchronized to the high speed of vehicle, and suit for different road conditions, complex and changing climatic. So difficulties and goals of the issue are the study of new machine vision algorithms for real-time and with good robustness. This thesis will focus on implementation of stereo matching algorithms for real-world images in DAS using GPU.Because of matching difficulties for real-world images, global algorithms are adopted in DAS. After introductions of preparatory works on stereo matching, such as camera calibration, stereo images rectification, common global algorithms and parallel programming using GPU, we focus on how to achieve efficient stereo matching algorithms (DP and BP) using CUDA for real-world images in DAS. Here is the major content of the paper:Firstly, dynamic programming algorithm for stereo matching is studied. To achieve satisfied disparity map for complex real-world images, a cost function calculation based on shift window and energy function based on temporal propagation are applied in our work, which can effectively improve the quality of the disparity map. Finally the algorithm is parallel implemented on CUDA.Secondly, belief propagation algorithm for stereo matching is also studied, which can obtain better disparity map than dynamic programming, meantime, but worse robustness and longer computing time. To address this issue, Sobel or Residual preprocessing on stereo images is used. Finally the algorithm is parallel implemented on CUDA.