Use of multiple difference-of-Gaussian filters to verify geometric models

Abstract

Experiments are reported with a program which recognizes London buses in busy street scenes. The program uses a 3D model of a bus. together with knowledge of camera and scene parameters, to compute the visual appearance of possible target events. In earlier work this took the form of 2D templates, which identify locations within the image where abrupt discontinuities of grey-value are expected. Each template corresponds to an hypothesis which is verified by measuring the agreement between the predicted features and the response of a Sobel filter to the image data. This paper reports an extension to the program which offers improved selectivity, and a more general-purpose method of linking predicted features to detected imageattributes. Multiple difference-of-Gaussian filters perform an initial object-independent analysis of the image to provide an intermediate domain in which object-dependent features, predicted by hypothesized instances of the model, are tested.