Algorithms for Automatic Sensor Placement to Acquire Complete and Accurate Information

Abstract

There is a need to develop strategies to automatically position mobile sensors to acquire data that contains accurate and complete information. This data is required to build models that are needed for navigation, manipulation, inspection, object recognition and verification. Systematic, random and exhaustive sensing techniques are inefficient. Automatic placement strategies can economize on the use of resources and maximize the quality and content of the information that is acquired. In this thesis methods are developed to generate the viewpoints (position and orientation) for a two dimensional laser range scanner and other sensors with a high angular field of view (>180o), a low angular resolution, and a limited range of detection can acquire accurate and complete information. A candidate viewpoint must satisfy a number of detectability constraints: the target feature must be visible and enclosed within the sensor? angular and range field of view. Two other constraints that affect the density and accuracy of the data must also be satisfied. The methods use descriptions of the sensor, the feature and analytic expressions of the constraints to generate a region in cartesian space that encloses the set of viewpoints that will satisfy each of the constraints. The intersection of the candidate viewpoints for all the constraints is the set of positions from where the sensor can acquire an accurate and complete image. To simplify the problem a two step process is adopted. First, a position from where the sensor can detect the feature (satisfying all constraints) is selected, and then the sensor is oriented so that the feature is in the sensor? field of view. When a sensor is not able to acquire accurate and complete information in a single image, multiple viewpoints are generated. To generate these viewpoints the feature is recursively subdivided until for each sub-feature there is at least one viewpoint from where the sensor can acquire accurate and complete information. The information contained in the images that are acquired from these viewpoints is sufficient to construct an accurate and complete description of the entire feature. The methods developed in this dissertation are able to generate viewpoints that are inside the cavity of a concave feature. The methods also apply to sensors with wide angular field of view, limited detection range, and low angular resolution.