Robust Algorithms for Object Localization

Object localization using sensed data features and corresponding model features is a fundamental problem in machine vision. We reformulate object localization as a least squares problem: the optimal pose estimate minimizes the squared error discrepancy between the sensed and predicted data. The resulting problem is non-linear and previous attempts to estimate the optimal pose using local methods such as gradient descent su er from local minima and, at times, return incorrect results. In this paper, we describe an exact, accurate and e cient algorithm based on resultants, linear algebra, and numerical analysis, for solving the nonlinear least squares problem associated with localizing two-dimensional objects given two-dimensional data. This work is aimed at tasks where the sensor features and the model features are of di erent types and where either the sensor features or model features are points. It is applicable to localizing modeled objects from image data, and estimates the pose...

• Type: Journal
• Year: 1998
• Where: IJCV
• Authors: Aaron S. Wallack, Dinesh Manocha
• Keywords: IJCV 1998, Optimal Pose, Optimal Pose Estimate, Squares Problem