Stereo matching by neural network that uses Sobel feature data

We develop a stereo vision system using Sobel training data and neural networks. Sobel operators are first used to extract features of intensity, variation, and orientation from stereo image pairs. These features are used to train a BP neural network in order to obtain an adaptive matcher. The trained BP matcher can generate an initial or primitive disparity map that provides necessary correlation or corresponding SSD (sum of squared differences) in area-based matching methods. Following the BP training, we propose a matching algorithm that is useful in iteratively updating the primitive disparity map. We show that this update algorithm can improve the quality of the disparity map significantly. At the final stage, several constraints such as epipolar line, ordering, geometric and local-support are added to further refine the map. The empirical results show the efficiency of the BP matcher and the validity of our matching algorithm