Model-based Vision for Traffic Scenes using the Ground-plane Constraint
Published in Real-time Computer Vision
Eds: Chris Brown & Dimitri Terzopoulos
Department of Computer Science
The University of Reading
An overview is given of a model-based vision system able to locate, recognise and track multiple vehicles in complex traffic scenes, from monocular cameras. Individual technical developments have been reported extensively in conference and research papers. The aim of this article is to summarise the various techniques, and draw them together into a description of the overall approach. It also discusses the main processing bottlenecks, and reports ongoing work which we believe will soon allow the system to run comfortably in real-time.
The current system already operates in near real-time with quite modest computing requirements. There are four main stages of processing:
Each of these processing stages depends crucially on the use of the ground-plane constraint (GPC) which stipulates that vehicles stand on a known planar surface - the road. The GPC makes computationally feasible the use of high-level reasoning about 3D vehicles, moving in a 3D scene, which greatly helps to solve many challenging tasks for the visual analysis of traffic.
Last Modified: 05:03pm GMT, November 12, 1996