A TRINOCULAR CENSUS-BASED STEREO VISION SYSTEM FOR REAL-TIME APPLICATIONS

摘要

This work presents a trinocular stereo vision system for real-time applications. The algorithm uses a sparsed census transform and is capable of processing trinocular configurations with baselines of different length. At first, a camera calibration tool for inline or L-shaped three-camera-configurations is introduced. For calculating the desired depth images, an existing census-based stereo matching system for binocular stereo vision is modified and extended to support the new trinocular camera configurations. Because of the extensive usage of Single Instruction, Multiple Data (SIMD) instruction sets and the support of state-of-the-art multi-core CPUs, dense depth maps can be produced at interactive frame rates, i.e. at least with several fps. A method of fusing the results of two stereo matching processes is developed in order to obtain depth maps, which are more dense and which contain less false positive matches. It was our intention to achieve reproducible data about the quality gain of the extension by one additional camera. We could use the Middlebury stereo vision evaluation for the 3-camera configurations and observed a reduction of false matches by up to 40% compared to the original binocular system. With a resolution of 640×480 and 80 disparities, frame rates of 12fps were achieved on state-of-the-art multi-core CPUs.