Managing sensor performance uncertainty in a multi-sensor robotic system

Many applications of robotics currently utilize sensors so that information obtained from the environment can be interpreted to help intelligently control the robot. Recently there has been great interest in the topic of sensor fusion, whereby information from multiple sensors can be combined to gain a better understanding of the environment. The increasing use of multiple sensors, however, also leads to the increasing possibility that the information gathered by some of these sensors will not agree. This disagreement may be due to differences in the environment but may also be due to the performance of the sensors themselves.; The problem addressed by this dissertation is the evaluation of the performance of the sensors themselves in a multi-sensor system. This is the complement of the problem addressed by sensor fusion, namely the use of multi-sensor systems to gain reliable information about the environment. A system which can determine the reliability of its sensors is more robust since it can wisely decide which sensors are most appropriate for a given task. Such a system can also determine whether sensor conflicts are the result of poorly performing sensors.; This dissertation contributes to the field of robotics in four areas: (1) Examines the importance of sensor confidence management as part a comprehensive solution to sensor integration--An area of research previously ignored or trivialized, this dissertation argues that sensor performance must be seriously considered in a multi-sensor system. (2) Introduces of the concept of activities as a method for determining performance--As a common basis for comparing all sensors in a robotic system, regardless of their type, the concept of activities is proposed. (3) Develops a computational model using activities--The theory proposed in this research is demonstrated by developing a computational model of the theory and constructing a prototype implementation of this model. (4) Extends Dempster-Shafer theory of uncertainty--A key part of the prototype implementation, Dempster-Shafer theory falls short of meeting all the requirements of the model. One aspect, the accumulation of uncertainty, is not covered by the theory. This dissertation suggests some methods for representing increasing uncertainty within the Dempster-Shafer theory framework.; The model and implementation were tested with eighteen experiments, conducted using data gathered from mobile robots at the University of the Pacific and the Colorado School of Mines. Both robots were equipped with ultrasonic transducers and a video camera, and data was gathered in a variety of different environments. The experiments demonstrate the feasibility of sensor performance measurement and the utility of activities for multi-sensor systems in unknown environments.