search and track, the interactions between the continuous planning framework and the underlying
sensing and control frameworks can be studied in
detail. On the other hand, it is not a good environment
for testing autonomy robustness; for this, a more realistic outdoor environment is required.
As noted, an executive dispatches plans and monitors their progress, possibly triggering new planning.
In search and track, the executive must be able
monitor the progress of the search for a target and to
switch to tracking mode if the target is found.
To test the executive in our controlled environment, a human target would begin out of sight of the
sUAV. The sUAV would take off and begin a pattern
maneuver (such as a square or spiral pattern) in search
mode. Once the human target was found, we studied
the response time of the system to transition into
track mode. Once in track mode, the target would first
move in a slow walk to demonstrate simple tracking
behavior. At some point, the target would take evasive
maneuvers until the subject was out of sight to the
sUAV. We then could observe the transition back into
search mode, which consisted of following another
pattern search (typically a simple rotation). This alternating behavior of search and track typically was
repeated many times in a single run.
Figure 6 shows a screenshot of the run time be-
havior of the system. The top left window shows
the output of the human detector window with the
bounding box. The bounding box is inherited by the
CMT tracker, as shown in the bottom right window.
The top right window shows the output of the IBVS
controller. The bottom left window shows part of the
SMACH planner visualization.
Figure 7 illustrates the behavior of the executive.
The variables person_detected and confidence in
the CMT tracker are monitored by the executive
to determine whether the system should be
searching or tracking. The system is designed to
switch from searching to tracking only when the
detector finds the human figure and the confi-
dence variable exceeds the desired threshold. This
aspect is important as both variables are prone to
noise because of their dependency on visual
Other indoor tests explored alternatives to target detection. First, a human detector model based
on training a deep neural network was used to
identify persons with backpacks (for example, to
be used by an sUAV to monitor for drug smuggling
across borders). Figure 8 shows a variation of the
indoor experiment in which the system distinguished a human being with a backpack from one
with no backpack. Other tests examined search
and track with a thermal camera. The system was
able to identify multiple human beings from only
a low- resolution thermal image using deep
Figure 8. Tracking Person with Backpack Only.