The authors present methods for refining sensor input from a sonar array by combining the information obtained from multiple readings. The sensors described in the paper are off the shelf components which measure with a rather wide, 30 degree, beam angle.  There are 24 sensors spaced at 15 degree intervals around a ring.

First, in order to combat noise and basic errors inherent in the sensor readings, they suggest techniques such as threshholding the data near the minimum and maximum of the sensor range, averaging of multiple readings, and clustering those readings to deal with situations with multiple obstacles along the same trajectory.  These approaches are used to obtain an overall sonar profile from a single robot location.

The readings are then mapped to probabilities.  If the sonar detects an obstacle at a range, $x$, then it is probable that the area closer than $x$ is empty but the area in the immediate vicinity of $x$ is occupied.  However, since the beam width is 30 degrees, sonar readings from a single robot position provide only a very vague sense of where an obstacle might be.  Thus, readings from multiple locations are combined probabalistically to refine the notion of whether an area is empty or occupied.

In order to combine readings, the individual maps must be oriented correctly relative to each other.  To this end, the authors also present an approach for determining the transformation needed to obtain an orientation match.  Essentially they iterate through possible transformations comparing occupied regions until a maximal similarity is found.  For the sake of efficiency, this is done initially at a very low resolution and then repeated at higher resolutions but with a more focused search of the transformation space.

Critique

Although the material presented in the paper is good, there are some issues that should have been discussed or explored in more detail.

The algorithms employed here make the assumption that the environment is fixed.  This is significant, as it is not an inherent requirement of sonar.  Furthermore, that assumption is not valid in many situations in which sonar is typically employed (ex. in submarines and as camera range finders).

Their choice of distribution functions for the probability that a region is empty or occupied are arbitrary and are not explained or justified.  Although their choice does not seem unreasonable, it would seem that the effectiveness of the overall algorithm would be highly dependent on these distributions.  Perhaps more attention should have been given to this aspect.

There is no quantitative analysis of the approach's effectiveness. For example, they could compute the percentage of ``false positives'' (empty areas marked as occupied) and ``false negatives'' (occupied areas marked as empty) or there could be some mathematical measurement of the map's accuracy, based on measured distance versus actual distance.  This computation would have been useful, for example, in evaluating the choice of probability distributions.

There is no discussion of how the position from which measurements is chosen.  Presumably certain locations, relative to obstacles, will result in more detailed measurements.  Quite likely the choice of position should be based upon the map, in order to refine areas with high uncertainty.