Multi-agent path planning for an unknown number of targets over dynamic space partitions

Searching a surveillance area for an unknown number of targets by making observations with small fields of view relative to the surveillance area is a task performed in settings such as search and rescue, area patrol, and reconnaissance. Based on observations, locations of possible target existence within the surveillance area can be estimated through time. This estimation can then be used to determine search paths. Typically this is done by optimization directly over the target estimation. Due to complicating factors such as the number of targets being unknown, this approach is not taken. Instead, time-evolving characteristic regions within the surveillance area are learned and paths are planned based on these regions. Based on the target estimation, three types of regions can be identified. These types are 1) regions of possible target existence, 2) regions of almost certain target existence, and 3) regions of almost certainly no targets. According to these regions, the surveillance area can then be partitioned into areas characteristic of required search or tracking and resource needed to perform the search or tracking. In order to accomplish this partitioning, a set of features is determined and used for classification. Once the surveillance area is partitioned, path planning is performed over the partitions. Because several partitions have been constructed, they can be allocated as tasks to a team of autonomous agents. The path planning is then separated into two levels. The first level is agent route planning over the partitions. The second level is agent path planning by optimization over partition level target estimation. Results are presented for a team of autonomous aerial vehicles searching for an unknown number of targets.