Underwater Volumetric Occupancy Mapping with Imaging Sonar

Underwater robotics is a strongly growing field with economic interests, such as in offshore industry, underwater archaeology and inspection tasks. One of the key contributors to more autonomy under water is the capability to build volumetric maps of the underwater world. Particularly for navigation, volumetric occupancy maps are one of the foci in research. The sensor imagery is often provided by forward scan sonars, which are mounted directly on underwater robotic platforms. Compared to light-based sensors, they handle turbid water significantly better and the low frequencies have less problems with attenuation than light waves. Existing approaches to build volumetric occupancy maps struggle to be cost efficient in terms of computation time or perform poorly in unknown environments. Additionally, the recovery of elevation information from forward scan sonars poses difficult challenges for map makers. This work introduces a novel sensor model for forward scan sonars that takes uncertainty in three dimensions into account. This sensor model is applied on preprocessed imagery and provides in three dimensional probability tensors. We show that the sensor model performs well in two and three dimensions and that elevation recovery through uncertainty quantification is a valid approach. Finally, the model is applied to imagery from a sonar simulator, where the integration of the model into Voxblox is verified