Uncertainty-Augmented Semantic 3D Map Labeling in Mobile Robotic Applications
Abstract: Autonomous navigation systems which operate in unknown or partially known environments strongly rely on sensor data to estimate the world state and do action planning. Considering such a scenario, the accuracy of the information supplied to the system is vital to make it able to operate correctly and safely. As a specific field of intelligent systems --in robotic applications-- it is desired to have semantic information of the environment for better recognition and navigation. For semantic perception, classification of objects into multiple classes is a fundamental requirement. On the other hand, it is an error-prone process as a consequence of variability of the scenes a robot may visit. However the lack of reliability of the classification step is often ignored or not explicitly measured. This problem is disregarded in existing literature. This master thesis work introduces a method to include uncertainty in the system. To this purpose, first the classifier and the specific dataset using an autonomous robot are created. Then the world scene is segmented and this information is projected to the 3D map. Later the uncertainty is estimated at different locations of the map. This map can then serve as input for variety of tasks such as exploration, active learning and human-robot interaction. Results on specific dataset generated using images from environment of TUM Theresienstrasse Campus show that uncertainty of classification is higher in misclassified areas.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)