Domestic service robots are becoming increasingly available and affordable. However, interaction with the real world is a challenge. Tracker systems are fast and can yield good result because they exploit temporal coherence, but often need manual initialization. Such a tracker system was combined with a detector system that initializes and corrects it automatically. Different (combinations of) features were explored and evaluated on a dataset. The combined system was also used to guide a mobile, autonomous robot in the real world. Data was gathered as the robot performed its task. This data was used to train the tracker system, to make it more robust to changes in environmental conditions such as illumination. This robustness is necessary if the tracking system has to handle the changing conditions in the real world without continuous feedback from an external system. The combined detection and tracker system was found to be able to run faster or perform better than the detector system in isolation, depending on the combined system's settings. Combination rules that adapt to the data helped to choose the best feature(s) depending on the conditions. In an evaluation on data gathered in the real world, the trained tracker performed better than the non-learning tracker.
Computer Vision, Robotics, Machine Learning
Herke van Hoof. (2011)
Interaction between Face Detection and Learning Tracking Systems for Autonomous Robots. Master Thesis. University of Groningen, Groningen, the Netherlands.