Multiresolution depth map estimation in PTZ camera network

Abstract

In this chapter, an active stereo vision system composed of two pan-tilt-zoom (PTZ) cameras is proposed for estimating multiresolution depth map for a large and complex scene. The rectification of stereo images is performed based on the sigmoid interpolation with a set of neural networks. The orientation parameters (pan and tilt values) and the rectification transformations of corresponding images are used as the input-output pairs for network training. The input data is read from cameras directly, whereas the output data is computed offline. The trained neural network is used to interpolate rectification transformations in real time for the stereo images captured at arbitrary pan and tilt settings. The correspondence between the stereo images is obtained using a chain of homographies based scheme. Non-homogeneity between the intrinsic parameters of two cameras is treated by means of zoom compensation to improve the quality of stereo rectification. Experimental results are given for estimating multiresolution depth map for a scene.