This thesis proposes an improved method of model building for mobile robot.
Multiple sensors are used to build model more consistently. We use stereo vision to detect vertical edges, and use ultrasonic sensors to detect vertical planes. By using multiple types of features, model building becomes less sensitive to the environment of robot. Difficulties in sensor fusion are solved partly by fusing sensor information on features to find symbolic entities.
Sensor uncertainty is reduced by fusing observations from multiple viewpoints. Computational cost for matching between corresponding features is greatly reduced by handling uncertainties at the model-level. Uncertainty of symbolic entity is computed from uncertainties of features comprising the entity. We use Kalman filter approach to reduce the uncertainty.