Demo Video
Summary
Global localization is localization with a given map, but does not use recent location records. These are simulations that estimate the current position based on a single camera input.Current Position
Estimated Position
Explored Path
Small Space
Large Space
Monte Carlo localization (MCL) is a widely used method for estimating the location of mobile robots. MCL works well in most environments. However, it does not work well with "kidnapped" problems, which have no previous consecutive sensor input. "Image retrieval" is effective in these "kidnapped" problems. "Retrieval" does not require consecutive sensor input, and "image" contains more information than other types of sensors. The problem with "Image" is that there is so much information that it is difficult to extract the information needed for our purpose. This problem can be solved using deep learning. Using Deep learning, we can extract the descriptor of the current camera image and compare it with the recorded descriptors in the database to find the most similar explored point. Through this, the current position can be estimated only by a single camera input.