An approach to model-based 3-D recognition of vehicles in real time by machine vision

An object recognition system for autonomous vehicle guidance applications is presented for interpreting monocular image sequences of motorway scenes in real time. The image sequences are taken by a camera mounted behind the windscreen of a test vehicle (VaMoRs) moving relative to the environment on a motorway. The approach described is based on an integrated spatio-temporal 4-D model for object tracking and relative state estimation and on knowledge based methods for feature matching and hypotheses testing. The recognition task is solved by combining a fast obstacle detection and classification algorithm using only 2-D image information with a module for 3-D object tracking. The vehicles are internally represented as 3-D polygonal models. They may be even recognized in the case of partial occlusion. The classification algorithm processes three different object classes: trucks or buses, vans and hatchback passenger cars. Within these classes the most relevant shape parameters are automatically adjusted during runtime. Up to two independently moving objects in the perceived scene can be processed simultaneously in real time, that means a cycle time of 200 ms. The system implemented on a transputer cluster has been tested by using noise corrupted measurements of synthetic and real images of German motorways