The problem of locating a robot within an environment is significant particularly in the context of mobile robot localization and navigation.; This thesis presents a new approach to mobile robot localization that avoids the selection of landmarks and the use of an explicit model. Instead, it uses the low level primitive features from video data, and learns to convert these features into a representation of the robot pose. The conversion from video data to robot poses is implemented using a multi-layer neural network trained by back-propagation. In addition, a key aspect of the approach is the use of the confidence measure to eliminate incorrect estimate components of pose vectors, and dead reckoning to complement the neural network estimates. Finally, the approach is generalized to allow a mobile robot navigate in a large environment.; Presenting a number of experimental results in several sample environments, the thesis suggests the accuracy of the technique is good while the on-line computational cost is very low. Thus, accurate localization of a mobile robot is achievable in real time.
展开▼
