M5.2 STEREO MATCHING USING A NEURAL NETWORK1
Yi-Tong Zhou, Rama Chellappa · 1988
A method for matching stereo images using a neural network is presented. Usually, the measurement primitives used for stereo matching are the intensity values, edges and linear features. Conventional methods based on such primitives suffer from amplitude bias, edge sparsity and noise distortion. We first fit a polynomial to find a smooth continuous intensity function in a window and estimate the first order intensity derivatives. A neural network is then employed to implement the matching procedure under the epipolar, photometric and smoothness constraints based on the estimated first order derivatives. Owing to the dense intensity derivatives a dense array of disparities are generated with only a few iterations. This method does not require surface interpolation. Computer simulations to demonstrate the efficacy of our method are presented. Stereo matching is a primary means for recovering 3-D depth from two images taken from different viewpoints. The two central problems in stereo matching are to match the corresponding points and to obtain a depth map or disparity values between these points. In this paper we present a method for computing the disparities between the corresponding points in two images recorded simultaneously from a pair of laterally displaced cameras based on the first order intensity derivatives. An implementation using a neural network is also given. Basically, there exist two types of stereo matching methods: region based and feature based methods according to the nature of the measured primitives. The region based methods use the intensity values as the measurement primitives. A correlation technique or some simple modification is applied to certain local region around the pixel to evaluate the quality of matching. The region based methods usually suffer from the problems due to lack of local structures in homogeneous regions, amplitude bias between the images and noise distortion. Recently, Barnard [l] applied a stochastic optimization approach for the stereo matching problem to overcome the difficulties due to homogeneous regions and noise distortion. Although this approch is different from the conventional region based methods, it still uses intensity values as the primitives with the aid of a smoothness constraint. Barnard’s approach has several advantages: simple, suitable for parallel processing and a dense disparity map output. However, too many iteations, a common problem with the simulated annealing algorithm, makes it unattractive. It also suffers from the problem of amplitude bias between the two images. The feature based methods use intensity edges or linear fea