The Central Complex Path Integration Circuit Deployed on an MAV
335
calculation. However, given the use of saccade mode, performance degradation
due to pitch angle increase in the flat and bumpy worlds is modest in comparison to the natural scene which features large asymmetric ramping sections that
cause large divergences between the assumed and actual depth values.
Height Compensation is Required for Successful Homing. Since neither
the use of height estimation or maintenance of a constant ground height is a given
for flying insects, we explored the impact of removing height input in the flat
world, while introducing a randomly varying height profile along the outbound
journey. Specifically, height profiles were generated by adding a random positive
offset to 10 equally spaced points along the outbound journey (z offset = U(0,
maximum z offset)), and then making a linear interpolation along this height
series according to the time into the outbound flight. The results, included in
Fig. 6e, indicate that even with a modest maximum height offset of 2 m, homing
performance is severely impacted if no height compensation is used in the visual
processing pipeline.
4 Discussion
We have implemented the CX model presented in [11] on an autonomous MAV.
This anatomically constrained neural circuit integrates ground speed in different
compass directions to maintain an estimate of location with respect to an origin,
providing a steering output to return home. Here we have introduced a biologically plausible matched filter to obtain ground speed from ventral optic flow. We
show that in real world testing, which includes asymmetric inbound/outbound
routes, random yaw saccades, and dynamic lighting and weather conditions, the
worst case final displacement error increase 1.5 m 100 m of outbound distance.
Thus this bio-mimetic system constitutes a medium range navigational aid with
comparable accuracy to a GPS module over the range of 1 km.
This accuracy could bring a flying insect near enough to its goal to find
it by search, but the decreased heading accuracy at short range (in Fig. 6b)
suggests that efficient relocation of the nest requires alternative mechanisms.
Moreover, some species of bee are known to commute several kilometers [3], so
it seems likely that additional guidance cues such as odour and vision are used
over longer ranges. In future work we will examine biologically plausible ways of
extending the navigational range and improving close range homing capability
of our biomimetic platform using drift free visual cues.
Our efforts primarily focused on developing a biologically plausible visual
ground speed estimation system, consisting of a matched filter based optical flow
subsystem. We have demonstrated that the acuity and frame rate of this system
could be as coarse as 3
◦ /pixel 10 Hz respectively. Best results were achieved with
1) the camera directed towards the ground, 2) mechanical stabilisation against
aircraft pitch and roll, and 3) height compensation. No discernible improvements
in homing performance were observed in relation to an increasing field of view
of the matched filter output, which could explain why neurons with a receptive
335
calculation. However, given the use of saccade mode, performance degradation
due to pitch angle increase in the flat and bumpy worlds is modest in comparison to the natural scene which features large asymmetric ramping sections that
cause large divergences between the assumed and actual depth values.
Height Compensation is Required for Successful Homing. Since neither
the use of height estimation or maintenance of a constant ground height is a given
for flying insects, we explored the impact of removing height input in the flat
world, while introducing a randomly varying height profile along the outbound
journey. Specifically, height profiles were generated by adding a random positive
offset to 10 equally spaced points along the outbound journey (z offset = U(0,
maximum z offset)), and then making a linear interpolation along this height
series according to the time into the outbound flight. The results, included in
Fig. 6e, indicate that even with a modest maximum height offset of 2 m, homing
performance is severely impacted if no height compensation is used in the visual
processing pipeline.
4 Discussion
We have implemented the CX model presented in [11] on an autonomous MAV.
This anatomically constrained neural circuit integrates ground speed in different
compass directions to maintain an estimate of location with respect to an origin,
providing a steering output to return home. Here we have introduced a biologically plausible matched filter to obtain ground speed from ventral optic flow. We
show that in real world testing, which includes asymmetric inbound/outbound
routes, random yaw saccades, and dynamic lighting and weather conditions, the
worst case final displacement error increase 1.5 m 100 m of outbound distance.
Thus this bio-mimetic system constitutes a medium range navigational aid with
comparable accuracy to a GPS module over the range of 1 km.
This accuracy could bring a flying insect near enough to its goal to find
it by search, but the decreased heading accuracy at short range (in Fig. 6b)
suggests that efficient relocation of the nest requires alternative mechanisms.
Moreover, some species of bee are known to commute several kilometers [3], so
it seems likely that additional guidance cues such as odour and vision are used
over longer ranges. In future work we will examine biologically plausible ways of
extending the navigational range and improving close range homing capability
of our biomimetic platform using drift free visual cues.
Our efforts primarily focused on developing a biologically plausible visual
ground speed estimation system, consisting of a matched filter based optical flow
subsystem. We have demonstrated that the acuity and frame rate of this system
could be as coarse as 3
◦ /pixel 10 Hz respectively. Best results were achieved with
1) the camera directed towards the ground, 2) mechanical stabilisation against
aircraft pitch and roll, and 3) height compensation. No discernible improvements
in homing performance were observed in relation to an increasing field of view
of the matched filter output, which could explain why neurons with a receptive