Efficient dearch and tracking for non-stationary targets

Abstract

Target search and tracking are elds which stand to bene t from guidance by intelligent, autonomous robots. We consider the problem of building an algo- rithm to guide a target search and tracking algorithm to autonomously search for, and track, a moving target. The thesis is motivated using the example of a drone which is tasked with nding and tracking a target in a restricted area. The drone is constrained by sparse signal information, time, and its limited knowledge of how the target moves. Most of the existing methods used to guide such robots, such as Bayesian Optimization [6][25] and Kinematic Motion Mod- els [4], do not consider both the search and tracking elements of the problem. The remaining methods in use, such as the Dirichlet-Multinomial pseudo-count [32][26], are unsuitable because they require a long training period before they are e ective. This thesis proposes a solution which uses elements of a Wonham Filter [33][12] to combine a proven search algorithm, Bayesian Optimization, with a ubiquitous tracking technique, Kinematic Motion Models. Secondarily, the solution considers a novel way of optimizing the search by using unsuc- cessful searches. The results show that the primary proposed solution may be more e cient at nding and tracking a target than existing methods, in certain circumstances.