Stereo Matching Under Speckle Noise

As human beings we are able to assess our surroundings through the process of stereo vision. Stereo vision involves taking our perceptions from our left and right view, and combining them to form our line of vision. Although this process is immediate and effortless for us, it can be quite complex when trying to apply it to computer vision. Stereo matching, which is responsible for finding the disparity map between two images to construct a 3D view of a scene, is a crucial step in stereo vision that is computationally very complex. The procedure involves finding the correspondence points between two images, which can be heavily affected by the quality of the input images they receive. The accuracy of the disparity map produced by this process becomes a factor of great importance when applied to different applications, i.e. military applications that involve automatic target detection or vision based obstacle avoidance. In this study, the performance of a traditional solution, the simulated annealing algorithm, is explored for radar images that have been introduced to speckle noise. From past experiments, it seems that the introduction of an NCC cost function has proven to aid simulated annealing in determining the correspondence points between image pairs. The cost function is controlled by two parameters, which heavily affect the accuracy of the produced disparity map. Hence, this study focuses on finding optimal values for those parameters when the radar images that contain speckle noise are used as input.