detector (Dalal and Triggs 2005) was integrated with
the trajectory planner to search for a human target.
Plan generation for search and track and the dispatching of plans were implemented using either
ROSPlan (Cashmore et al. 2015) or SMACH (Bohren
et al. 2011) for comparison. A Parrot AR.Drone quad-rotor was commanded from a computer via WiFi link
using the AR.Drone Autonomy ROS package2 or in
simulation using Gazebo (Koenig and Howard 2004) as
well as MORSE-Blender. All simulations were run under
Ubuntu 14.04 LTS 64-bit and an Intel Xeon E5-2630 @
2. 60 GHz × 17 CPU, a NVIDIA Quadro K5000 GPU,
and 32 GB of RAM.
The simulation platform provided safe and fast
opportunities for comparing the performance of
different algorithms or models (figure 4) and
allowed a quick way to fine-tune certain parameters
such as controller gains and delays between target
identification before flight experiments. Other
trade studies for autonomy that benefitted from
simulation include the choice of onboard sensors
and the study of communication delay. For example, thermal cameras were studied for nighttime
For target search from an sUAV, one important
parameter is identification of the target from a moving
platform. Images from a moving vehicle are often
subject to motion blur. This makes object identifi
cation extremely difficult. Figure 5 shows how motion
blur was reproduced in simulation. Simulated tests
allowed for the discovery of the maximum velocity of
the vehicle in which object detection still worked.
Furthermore, it was found that different algorithms
for object detection had different effects on the blur.
While deep learning methods always outperformed
simple HOG-Haar networks for pedestrian detection,
they were also more susceptible to missed detection in
case of motion blur. These factors helped us select the
maximum speed at which the vehicle was allowed to
operate while searching for a target.
In addition to simulation, flight tests during conceptual design were conducted in an indoor test facility at the NASA Ames Research Center. Some of the
results summarized here are discussed in more detail
elsewhere (Chakrabarty et al. 2017). The indoor facility
at the Ames Research Center is a controlled environment that can be used to test systems before moving
on to realistic outdoor environments. Specifically, for
Figure 6. Testing Autonomous Search and Track at the Ames Research Center Indoor Facility.