Research On Stereo Matching Hardware Architecture Based On Belief Propagation

Stereo vision,as one of the important hot topic in computer vision,has a broad application prospect in many areas such as target recognition,automatic navigation and industrial control automation.Stereo matching is one of the key component of the stereo vision system.The precision of stereo matching determines the accuracy of subsequent depth acquisition and 3D reconstruction,meanwhile its operation speed affects the real-time characteristics of the whole stereo vision system.As a global stereo matching algorithm,the belief propagation performances better than most of the local ones and the similar operation between nodes provides us the possibility for hardware acceleration.However,the constant iteration in the belief propagation results in a large amount of hardware resource consumption and computational delay.Therefore,this thesis focus on the low complexity,real-time BP algorithm of hardware realization in stereo matching.First,this thesis reviews the principle of stereo vision system,introduces the background and basic algorithm of stereo matching and the evaluation methodology.In this thesis,the global energy minimization method is chosen as the evaluation methodology of optimal performance.Then,based on the BP algorithm,this thesis compares the two basic message passing schemes in the ideal hardware environment and analyzes the challenges in hardware implementation.In addition,the BP-M with better optimization performance is chosen as the optimization performance benchmark.Secondly,this thesis introduces two improved algorithms of belief propagation.The key problems in hardware implementation are analyzed and compared,such as the the real-time bottleneck of the tile-based BP algorithm and the influence of the number of levels and interations on the performance of hierarchical BP.An improved real-time scheme based on Tile-based BP algorithm is proposed for image with low resolution.Finally,the real-time performance of the stereo matching algorithm based on belief propagation is improved from two aspects.In the aspect of algorithm,two real-time combinatorial algorithms for hardware realization based on hierarchical BP and tile-based BP are proposed.It is found that both of these two algorithms can achieve or even exceed the traditional BP-M in terms of optimization performance,moreover,the proposed algorithm which based on Tile-based BP has a great improvement in the real-time characteristics,and hardware resource consumption is controlled within the reachable range.In the aspect of processing element,this thesis designs and implements two processing elements,including a processing element based on the traditional min-convolution algorithm,and a low-complexity processing element.Then,behavioral simulation and logical synthesis are proformed for both the two kinds of processing element.By comparing the synthesis results,the low-complexity processing element has higher hardware efficiency with a lower delay,which can meet the requirements of the proposed combinatorial algorithm and enhance the real-time performance of the system.