A High-Performance Vision System for Obstacle Detection

Abstract

I have developed a method for obstacle detection in a highway environment using the CMU Video Rate Multi-Baseline Stereo Machine. One key feature of this method is the computation and output of a confidence measure at each pixel, so that regions of the scene with low reliability can be recognized and filtered out. This method has proven to be capable of detecting 30cm high obstacles at distances of close to 100m. Unfortunately, the system requires significant of post-processing of the output data from the CMU stereo machine, and therefore cannot be run in real-time. Additionally, the design of the CMU stereo machine imposes several limitations on the system, both computationally (limited resolution, speed, and programmability) and practically (large size, cost, power consumption). I propose to construct a system, consisting of both hardware and software, to perform this obstacle detection task in real-time. In the process, I will develop a general-purpose stereo vision machine that improves upon the state-of-the-art. Furthermore, I will show that the system can perfom obstacle detection in real-world environments and thus improve vehicle safety.