next up previous
Next: Grid-based Representation of the Up: Filtering Techniques for Dynamic Previous: The Entropy Filter

The Distance Filter

 

The distance filter has specifically been designed for proximity sensors such as laser range-finders. Distance filters are based on a simple observation: In proximity sensing, unmodeled obstacles typically produce readings that are shorter than the distance expected from the map. In essence, the distance filter selects sensor readings based on their distance relative to the distance to the closest obstacle in the map.

To be more specific, this filter removes those sensor measurements s which with probability higher than tex2html_wrap_inline3169 (this threshold is set to 0.99 in all experiments) are shorter than expected, and which therefore are caused by an unmodeled object (e.g. a person).

To see, let tex2html_wrap_inline2987 be a discrete set of possible distances measured by a proximity sensor. As in Section 3.2, we denote by tex2html_wrap_inline3175 the probability of measuring distance tex2html_wrap_inline3001 if the robot is at position l and the sensor detects the closest obstacle in the map along the sensing direction. The distribution tex2html_wrap_inline3043 describes the sensor measurement expected from the map. As described above, this distribution is assumed to be Gaussian with mean at the distance tex2html_wrap_inline2985 to the closest obstacle along the sensing direction. The dashed line in Figure 9 represents tex2html_wrap_inline3043 , for a laser range-finder and a distance tex2html_wrap_inline2985 of 230cm. We now can define the probability tex2html_wrap_inline3189 that a measured distance tex2html_wrap_inline3001 is shorter than the expected one given the robot is at position l. This probability is obviously equivalent to the probability that the expected measurement tex2html_wrap_inline2985 is longer than tex2html_wrap_inline3001 given the robot is at location l and thus can be computed as follows:

  eqnarray848


  figure1231

In practice, however, we are interested in the probability tex2html_wrap_inline3225 that tex2html_wrap_inline3001 is shorter than expected, given the complete current belief of the robot. Thus, we have to average over all possible positions of the robot:

  eqnarray863

Given the distribution tex2html_wrap_inline3225 , we now can implement the distance filter by excluding all sensor measurements tex2html_wrap_inline3001 with tex2html_wrap_inline3233 . Whereas the entropy filter filters measurements according to their effect on the belief state of the robot the distance filter selects measurements solely based on their value and regardless of their effect on the robot's certainty.

It should be noted that [Fox1998] additionally developed a blockage filter for proximity sensors, which is based on a probabilistic description of situations in which a sensor is blocked by an unknown obstacle. We omit this filter here since its derivation is quite complex and the resulting filter is not significantly different from the distance filter described here.


next up previous
Next: Grid-based Representation of the Up: Filtering Techniques for Dynamic Previous: The Entropy Filter

Dieter Fox
Fri Nov 19 14:29:33 MET 1999