Gaussian Gridmaps from Gaussian Processes for WiFi-Based Robot Self-Localization in Outdoor Environments

Gaussian Processes have been effectively used to learn location-to-signal-strength mappings from previously acquired observations and enable WiFi-based robot self- localization. However, the cubic computational cost for training and the quadratic cost for prediction with respect to the number of training points limits their scalability, particularly with large datasets necessary for outdoor environments. To reduce prediction cost we propose the use of Gaussian Gridmaps, a spatial representation that stores mean and variance predictions from Gaussian Processes into gridmaps. This approach reduces prediction computational cost to constant time, at the expense of some localization accuracy and increased memory usage. Our experiments demonstrate the feasibility of this method for outdoor localization and examine the impact of quantization and grid resolution on localization performance.