Stereo-Vision-Support for Intelligent Vehicles - The Need for Quantified Evidence

Abstract

Vision-based driver assistance in modern cars has to perform automated real-time understanding or modeling of traffic environments based on multiple sensor inputs, using ‘normal’ or specialized (such as night vision) stereo cameras as default input devices. Distance measurement, lane-departure warning, traffic sign recognition, or trajectory calculation are examples of current developments in the field, contributing to the design of intelligent vehicles.

The considered application scenario is as follows: two or more cameras are installed in a vehicle (typically a car, but possibly also a boat, a wheelchair, a forklift, and so forth), and the operation of this vehicle (by a driver) is supported by analyzing in real-time video sequences recorded by those cameras. Possibly, further sensor data (e.g., GPS, radar) are also analyzed in an integrated system.

Performance evaluation is of eminent importance in car production. Crash tests follow international standards, defining exactly conditions under which a test has to take place. Camera technology became recently an integral part of modern cars. In consequence, perfectly specified and standardized tests (‘camera crash tests’) are needed very soon for the international car industry to identify parameters of stereo or motion analysis, or of further vision-based components.

This paper reports about current performance evaluation activities in the .enpeda.. project at The University of Auckland. Test data are so far rectified stereo sequences (provided by Daimler A.G., Germany, in 2007), and stereo sequences recorded with a test vehicle on New Zealand’s roads.