334
J. Stankiewicz and B. Webb
environmental mesh. When the mesh was reskinned with a uniformly textured
pattern, performance is similar to the other non-uniform height worlds. The
forested world was the worst performing textured world, suggesting high depth
discontinuities are more problematic than smooth slopes. It should be noted that
in order to achieve this performance in the forest world, the altitude control had
to be adapted to maintain a fixed height above the ground, rather than using
range sensor readings. Insects sometimes fly over gaps without adjusting their
height but the ability to maintain a fixed height relative to the ground while
flying over a tree canopy is speculation.
Directional Certainty Increases with Distance from the Origin. Previous work predicts that while the positional uncertainty of a noisy PI circuit
increases with outbound foraging range, directional certainty increases [15] The
first of these assertions is supported by Fig. 6a and in Fig. 6b the second is
apparent. Here the moving average mean heading error is plotted for all range
ordered samples of the real dataset as a function of radial distance from the trial
origin. From geometric considerations the heading error should be ∝ positional
error/range [15]; hence inbound heading error will be lower for shorter outbound
distances when less positional error has accumulated (50 m curve in Fig. 6b) but
it is evident that for all distances, the inbound heading error increases rapidly
from under 10 m. This suggests accurate relocation of the exact origin would
require an alternative “drift free” cue such as local visual information.
A Reduced Field of View Does Not Affect Homing Performance.
We investigated whether the size of the field of view used to estimate ground
speed would affect the accuracy of homing performance, by cropping the input
before the final summing step in the visual pipeline. In Fig. 6c it can be seen
that there is no tangible impact for any of the tested environments, even when
the speed cell ground speed estimate is based on a visual field of only 2 × 2
pixels. An expanded field of view could potentially be detrimental given the
proportional relationship between depth and optic flow, as in a natural scene
this would increase the likelihood of encountering non-ground-plane objects in
the visual field. Conversely, wider field of view measurements could provide a
filtering and/or voting mechanism to reject flow field outliers.
Ventral Views Provide More Robust Optic Flow Information. In order
to understand the practical implications of sampling optic flow from different
view elevations, the view direction of the vision sensor was systematically raised.
In simulation, the single downward facing camera was replaced with a pair
of time synchronised cameras that were rotated −45
◦ and 45
◦ in yaw respectively. These cameras were then pitched equally upwards by 0,15,30,50 or 70
◦ .
It was necessary to use saccade mode (see Sect. 2.4) for these trials to prevent
a catastrophic reduction in homing performance in the Furkastrasse retextured
world. The results in Fig. 6d show that increasing the pitch angle reduces homing performance in all cases to some extent. This can be explained by the
greater variance of μ weighting factors for higher pitch angles (see Fig. 2) resulting in those areas with higher weighting contributing more noise to the speed
J. Stankiewicz and B. Webb
environmental mesh. When the mesh was reskinned with a uniformly textured
pattern, performance is similar to the other non-uniform height worlds. The
forested world was the worst performing textured world, suggesting high depth
discontinuities are more problematic than smooth slopes. It should be noted that
in order to achieve this performance in the forest world, the altitude control had
to be adapted to maintain a fixed height above the ground, rather than using
range sensor readings. Insects sometimes fly over gaps without adjusting their
height but the ability to maintain a fixed height relative to the ground while
flying over a tree canopy is speculation.
Directional Certainty Increases with Distance from the Origin. Previous work predicts that while the positional uncertainty of a noisy PI circuit
increases with outbound foraging range, directional certainty increases [15] The
first of these assertions is supported by Fig. 6a and in Fig. 6b the second is
apparent. Here the moving average mean heading error is plotted for all range
ordered samples of the real dataset as a function of radial distance from the trial
origin. From geometric considerations the heading error should be ∝ positional
error/range [15]; hence inbound heading error will be lower for shorter outbound
distances when less positional error has accumulated (50 m curve in Fig. 6b) but
it is evident that for all distances, the inbound heading error increases rapidly
from under 10 m. This suggests accurate relocation of the exact origin would
require an alternative “drift free” cue such as local visual information.
A Reduced Field of View Does Not Affect Homing Performance.
We investigated whether the size of the field of view used to estimate ground
speed would affect the accuracy of homing performance, by cropping the input
before the final summing step in the visual pipeline. In Fig. 6c it can be seen
that there is no tangible impact for any of the tested environments, even when
the speed cell ground speed estimate is based on a visual field of only 2 × 2
pixels. An expanded field of view could potentially be detrimental given the
proportional relationship between depth and optic flow, as in a natural scene
this would increase the likelihood of encountering non-ground-plane objects in
the visual field. Conversely, wider field of view measurements could provide a
filtering and/or voting mechanism to reject flow field outliers.
Ventral Views Provide More Robust Optic Flow Information. In order
to understand the practical implications of sampling optic flow from different
view elevations, the view direction of the vision sensor was systematically raised.
In simulation, the single downward facing camera was replaced with a pair
of time synchronised cameras that were rotated −45
◦ and 45
◦ in yaw respectively. These cameras were then pitched equally upwards by 0,15,30,50 or 70
◦ .
It was necessary to use saccade mode (see Sect. 2.4) for these trials to prevent
a catastrophic reduction in homing performance in the Furkastrasse retextured
world. The results in Fig. 6d show that increasing the pitch angle reduces homing performance in all cases to some extent. This can be explained by the
greater variance of μ weighting factors for higher pitch angles (see Fig. 2) resulting in those areas with higher weighting contributing more noise to the speed