A Probabilistic Approach to Autonomous Path Planning for Directional Mobile Sensors

Abstract

Directional sensors, such as vision, infrared, ultrasound, and active acoustic sensors, are characterized by a preferred sensing direction, such that measurements are obtained only for a bounded subset of all possible aspect angles. By such approach, directional sensors can obtain information about the target's relative orientation, in addition to its distance. Common applications include cameras mounted on autonomous vehicles that may be used for urban surveillance or target recognition by means of on-board computer vision algorithms. One of the major challenges in planning the motion of directional mobile sensors is that an important target of interest may be occluded by the presence of obstacles in the sensor's line-of-sight. This thesis addresses this path-planning problem for an Unmanned Ground Vehicle (UGV) equipped with a vision sensor for the purpose of classifying multiple static targets in an obstacle-populated environment. An approach is developed for determining a UGV path that enables observations from all targets with known locations in minimum time. The approach guarantees that the UGV is able to classify every target previously localized, while avoiding collisions with obstacles and occlusions that prevent line-of-sight visibility. The approach consists of mapping targets into the UGV configuration space, thus obtaining C-targets, using complexity reduction techniques that take into account shadow regions caused by the presence of obstacles. An information roadmap method (IRM) algorithm is used to build a connectivity graph from the C-target regions, and a solution with the least translation distance is obtained. Comprehensive simulations performed in MATLAB and Webots - a professional robot simulator that provides modules for sensors, robots, and their interactions with a 3-D virtual environment - demonstrate the effectiveness and performance improvement of the proposed approach when compared to existing methods, based on "nearest neighbor" and classical traveling salesman problem (TSP) algorithms.