Sensors are a very important component of any autonomous system. Mobile robots can use a variety of sensors to gather information about the current state of the environment they are operating in. Some of these sensors include infra red proximity sensors, sonar range sensors, cameras, audio sensors, gyroscopes, GPS, odometry sensors, laser range scanners, stereo vision camera systems etc. There are two major problems in attempting to fuse this data from multiple sources into a unified piece of useful information.
Firstly, the sensory data can take various forms: distances from range sensors such as sonar and stereo vision, visual images from cameras, audio signals from microphone, binary signals from proximity sensors such as infra-red detectors etc. So a common representation for all the sensory data is needed before we can attempt to fuse it. Secondly, the sensory data may be corroborative or contradictory. As an example, a camera image may show a car but the microphone may or may not predict the car depending on whether the engine is running or stopped. We need to be able to determine what relationships can give rise to such scenarios and model them.
In order to solve the problem of common representation, we generate probabilistic models of the robot sensors. These probabilistic sensor models, termed belief sensors convert sensory data into probabilistic data that can then be fused. Solving the problem of conflicting sensory data is much harder. Our approach here is to use expert knowledge (like in the car example above) to create confidences in our various sensors and then fuse the data accordingly. Thus, the audio sensor would get a lower confidence and we would accept the presence of a car even when only the visual sensor indicated it. However, the audio sensor can increase our belief in the presence of the car if both it and the visual sensor suggest it.
Our mobile robots are currently equipped with 2 bump sensors, a sonar range sensor, a camera, 12 infra red proximity sensors and encoders on the wheel motors. Chris Eveland is working on installing a stereo vision camera module on the robots. Craig Harman and Amit Singhal developed a client server interface to use Randal Nelson's object recognition system on the robots. Amit Singhal and Henry McCauley built a dead reckoning system that uses the wheel encoders to perform robot localization in a 2-D environment. In addition to modelling the sensors currently installed on the mobile robots, we are also building sensor models for other sensors to be used in simulation (and possibly installed later on the wheelchair). These sensor models include audio sensors and GPS.
This page is maintained by Amit Singhal.
Last update: 11/17/97.