Occupancy grid artefact removal and error correction using GANs

Occupancy Grid Mapping is a form of Simultaneous Localisation and Mapping (SLAM) in which the world around a robot is visually represented as a grid map. This form of map can be compared to a floor plan in which features within an environment such as walls are labelled in place. Certain issues such as noise, artefacts, linear error, angular error, and incomplete rooms make this representation difficult to appropriate. Generative Adversarial Networks (GAN) [1] in the past have proven successful in and reliable methods for noise reduction, artefact removal [2], and partial observation completion [3]. We demonstrate a novel data creation process to mass produce samples of erroneous and ideal occupancy grid maps. We use this data to build two GAN models based on well-known frameworks CycleGAN [4] and CUT [5] for the task of occupancy grid cleaning. We demonstrate the generalisability of our models through making predictions of ‘clean’ maps on samples of real data from the Radish Dataset [6].