EXPLORAÇÃO E BUSCA VISUAL ROBÓTICA EM AMBIENTE SIMULADO

In this work, a robotic agent named Angela has been devised and simulated. The virtual robot has two degrees of freedom and is capable of performing search and visual exploration in a virtual environment. Its modular architecture allows for straightforward substitution and improvements of its modules, aiming for better overall performance. During exploration, the visual attention module detects salient objects and triggers online learning; when in search mode, candidate objects are searched for by the object proposal generation module. The classification module, which determines whether the visualized object corresponds to the target, is comprised of oiSGNG (Online Incremental Supervised Growing Neural Gas), an original classification algorithm devised and implemented during this work. oiSGNG has been featured in IJCNN 2017 (International Joint Conference on Neural Networks). The image segmentation module is another contribution of this work: StochGrow, which is based on region growing and provides a solid solution for the tradeoff between speed and quality of real time image segmentation. The complete system can be tweaked to address several real world robotic problems.