The Central Complex Path Integration Circuit Deployed on an MAV
327
handled using MAVROS. The CX module was wrapped in a ROS node and
acted as a state in our generic behavioural state machine. A separate ROS node
was developed to handle ground speed estimation. A major benefit of using
the opensource PX4 ecosystem is that a customisable simulation model was
available for use in the GAZEBO environment. This provided a realistic physical
and graphical simulation setup, supporting rapid development and evaluation in
otherwise untestable configurations. The PX4 flight stack features an extended
Kalman filter software module (EKF2) which is used to provide ground truth
location estimates. The differential GPS system that supports EKF2 in our setup
has a reported horizontal accuracy of ±6 cm, the heading error of the ground
truth is unspecified.
The primary function of the camera in this system is to facilitate ground
speed estimation. We therefore opted for a downward facing configuration on
both the real and simulated MAVs. Insects are known to stabilise their head
during flight [14] and because MAVs have a high degree of body vibration and the
body necessarily pitches and rolls to generate translational thrust, we mimicked
head stabilisation with a 2-axis gimbal and found this essential for good ground
speed estimates (see Fig. 5).
2.2 Ground Speed Estimates from Optical Flow and Matched
Filters
Flying insects have been shown to estimate distance travelled from optic flow
[10]. Recent neural recordings in 2 species of bee have revealed two pairs of
motion sensitive neurons that respond most strongly to flow-fields corresponding
to animal head motion along orthogonal axes left (−45
◦ ) and right (45
◦ ) in
azimuth with respect to the animal’s anterior [11] (labelled TN left/right in
Fig. 1). We mimic these neural properties, but rather than replicate the nearpanoramic vision system of an insect, we used a conventional camera with a
42
◦ field of view, pointing downward. One advantage of this configuration is
that optical flow induced by pure yaw motion sums to zero when combined
with translational matched filters with preferred directions that are parallel to
the ground. However, as it is not apparent that the insect visual field actually
includes the area directly below the animal [13], we also investigate the effect of
raising the camera view angle in Sect. 3.
The processing pipeline for each speed cell is presented in Fig. 2a. The camera
operates at 10 Hz. The preprocessing stage converts to grayscale and resizes to
150 × 235 pixels. A 2-frame Farneb¨ ack dense optic flow algorithm [1] is used to
compute the optical flow field. This is combined with matched filters for selfmotion and weighted by a depth estimate for each pixel to obtain a ground speed
estimate as follows:
327
handled using MAVROS. The CX module was wrapped in a ROS node and
acted as a state in our generic behavioural state machine. A separate ROS node
was developed to handle ground speed estimation. A major benefit of using
the opensource PX4 ecosystem is that a customisable simulation model was
available for use in the GAZEBO environment. This provided a realistic physical
and graphical simulation setup, supporting rapid development and evaluation in
otherwise untestable configurations. The PX4 flight stack features an extended
Kalman filter software module (EKF2) which is used to provide ground truth
location estimates. The differential GPS system that supports EKF2 in our setup
has a reported horizontal accuracy of ±6 cm, the heading error of the ground
truth is unspecified.
The primary function of the camera in this system is to facilitate ground
speed estimation. We therefore opted for a downward facing configuration on
both the real and simulated MAVs. Insects are known to stabilise their head
during flight [14] and because MAVs have a high degree of body vibration and the
body necessarily pitches and rolls to generate translational thrust, we mimicked
head stabilisation with a 2-axis gimbal and found this essential for good ground
speed estimates (see Fig. 5).
2.2 Ground Speed Estimates from Optical Flow and Matched
Filters
Flying insects have been shown to estimate distance travelled from optic flow
[10]. Recent neural recordings in 2 species of bee have revealed two pairs of
motion sensitive neurons that respond most strongly to flow-fields corresponding
to animal head motion along orthogonal axes left (−45
◦ ) and right (45
◦ ) in
azimuth with respect to the animal’s anterior [11] (labelled TN left/right in
Fig. 1). We mimic these neural properties, but rather than replicate the nearpanoramic vision system of an insect, we used a conventional camera with a
42
◦ field of view, pointing downward. One advantage of this configuration is
that optical flow induced by pure yaw motion sums to zero when combined
with translational matched filters with preferred directions that are parallel to
the ground. However, as it is not apparent that the insect visual field actually
includes the area directly below the animal [13], we also investigate the effect of
raising the camera view angle in Sect. 3.
The processing pipeline for each speed cell is presented in Fig. 2a. The camera
operates at 10 Hz. The preprocessing stage converts to grayscale and resizes to
150 × 235 pixels. A 2-frame Farneb¨ ack dense optic flow algorithm [1] is used to
compute the optical flow field. This is combined with matched filters for selfmotion and weighted by a depth estimate for each pixel to obtain a ground speed
estimate as follows: