Optical Linear Feature Detection Based on Model Pose.
Mark R. Stevens, J. Ross Beveridge · 1995
Low-level edge detection in optical imagery can be problematic in the ATR domain where highly complex scenes are the norm. Feature detection algorithms typically take a global approach, resulting in the discovery of many fragmented lines which are not directly related to stored model information. For this domain, we have taken a top-down approach which searches an optical image for the locally optimal features based on the current hypothesized object pose. The resulting linear features can then be matched against a CAD model. 1 Introduction Edge detection in the Automatic Target Recognition (ATR) domain should be driven by the expectation of which model features are assumed to be visible in a given image. Using a hypothesized model pose to predict visible features from a CAD model [Mar96, Ste95], a local optimization procedure is used to find the corresponding and consistent data features in the image. The process differs from the traditional low-level bottom-up edge detection process...