Wavelet Occupancy Grids: A Method for Compact Map Building

The capacity to know about the environment is a major requirement for robots and automated systems. Real scenes are complex to perceive, dynamic and large. Therefore to perceive those scenes, robots have access to complementary sensors such as sonar/laser range-finders, cameras or bumpers, but all these are always noisy. In automated navigation, the internal representation of the environment is used for all fundamental tasks: localization[14], pathplanning[6], obstacle avoidance[1] or target tracking[2]. But this modelling can require huge memory space, especially when the level of abstraction is fairly low and the representation is close to the sensor data. In these cases the time of data analysis is also prohibitive, making it necessary to change sophisticated algorithms for naive ones. This paper addresses the problem of data representation and data storage for large maps, under the constraints of multi-sensor real-time updates and hierarchical representation. Thus multiscale algorithms could be applied to very large maps with efficient calculating time. The first step when mapping the environment is obviously to define how to map. Geometrical properties are powerful guidelines to do it. A focus on the ability to reach a point from an other leads to the class of topological maps and the associated representation is the graph of connected places. Topological maps are semantically rich, can be drawn even without an accurate localisation of the robot and their size is very small. However it is difficult to build and update them automatically and they do not allow direct complex data manipulations. The possibility to index all the spatial information in a common reference frame, via a global coordinate system, leads to the class of metric maps. In an inversely symmetrical manner, metric maps are easy