As described in Section 4.1, dense depth images are regularly constructed by the trinocular stereo vision service on-board the robot. These depth images can be reduced to represent the nearest obstacles by projecting all sensed points down through a vertical column to the plane of the floor. The depth image is reduced to a single row of disparities representing the closest obstacle as seen from a top-view perspective. This 2-D map has high angular resolution; however, range uncertainty varies proportionally with the depth. This representation is much smaller than full depth images and are much cheaper to send to the host.
In the host, the radial depth maps are routed by the mailer to the mapper module. The mapper application integrates these directional range maps into a 2-D map represented by an occupancy grid [7]. Such a map is represented by a tessellation of the mapped space into a grid. The value of each grid is related to the probability that this space is occupied by any part of an obstacle. The ``mapper'' initializes the map to contain only values at 50% probability, indicating that the entire space is unknown. As new range maps arrive, the mapper updates the occupancy grid so that each cell contains an updated probability that the cell is occupied by an object. Every point between the current position of the robot and the nearest obstacle in a given direction is marked clear. The probability of the cell at the given range is updated, combining its previous value with the uncertainty of the range estimate. Cells beyond the object detected are unaffected.
In a sense, the mapper acts as a smart memory, that integrates the information over time into a coherent whole, and buffers data between the vision service and the client.