Design And Implementation Of Binocular Vision System Based On Heterogeneous Multi-core Processor

As an important branch of computer vision,binocular vision can restore depth information by calculating the parallax of a space point in the same scene in binocular image,which has two advantages of non-contact and passive.In recent years,with the in-depth study of binocular matching algorithm,some algorithms have improved greatly in terms of accuracy,but they are all premised on the ideal binocular image.In fact,because of the influence of realistic imaging factors,it often results in the deviation of the binocular image in brightness and so on,which requires that the algorithm should have high robustness.In addition,the algorithms with high accuracy often have large computation and poor real-time performance,so it is difficult to apply those algorithms in the application of real-time processing.The SGBM(Semi-global block matching)algorithm is a semi-global matching algorithm with high accuracy,but its robustness and real-time performance still need to be improved.Therefore,this thesis made an improved research and a parallel acceleration method for the SGBM algorithm.And the improved algorithm is applied to the binocular ranging system in this thesis.Aiming at the problem that the BT matching cost in the SGBM algorithm is not robust to the luminance deviation,we have proposed a BT-Census matching cost of combining Census matching cost.Since the Census matching cost preserves the structural characteristics between neighborhood pixels,the robustness to the brightness deviation is improved during matching.In order to improve the real-time performance,we studied a parallel acceleration method of the algorithm,and we implemented the algorithm based on OpenCL heterogeneous parallel acceleration framework.Firstly,we made a parallel analysis of the main algorithms of matching cost,cost optimization,parallax calculation and parallax refinement modules in this algorithm.Then,based on OpenCL framework,we designed the data storage structure and OpenCL kernel of the algorithm.Finally,combined with kernel performance evaluation and algorithm optimization,we implemented this heterogeneous parallel acceleration algorithm.Compared with the SGBM algorithm,experiments have shown that the parallel acceleration algorithm in this thesis have 2.2 times speedup in real-time performance,with similar accuracy on the AMD APU A8-4555M processor.Based on the above improved algorithm,this thesis designed and implemented a binocular range system.This binocular range system used a heterogeneous multi-core computing platform based on APU,which not only reduced the data transmission delay between multiple cores,but also effectively controlled the system power consumption.Experiments have shown that the average ranging period of this binocular range system was 110 millisecond and depth measurement error less than 9.6 percent in the range of 0.5 to 5 meters for 640 x 480 pixel images with 64 disparity levels.This thesis initially achieved a low power,near real-time binocular range system.