Appearance-Based Robot Self-Localization

Robot Self-Localization is one of the key tasks in the robotics. There are different approaches to solving this problem. One possible approach is self-localization using visual data, the data that can be obtained with digital camera. In this work I introduce the algorithm of creating the dataset of Google Street View panoramas via Google Street View Api. The algorithm can be used to create big datasets for the appearance-based localization task. Using this algorithm I collected the dataset of 1518 panoramas, that were published by Google Street View for 1.5x2 km region. Each panorama is represented by 4 different images which correspond to 4 directions (north, south, east, west). In this work I also describe the neural network model based on pretrained SqueezeNet network and fine-tuned on the collected data. The final network successfully solves localization task and slightly outperforms the baseline nearest neighbour algorithm.