254
S. Meyer et al.
indication that the skyline height without further processing is not a robust
feature for orientation recovery at least in this simulated world.
All models show a noticeable amount of data points outside the 1.5 IQR
interval. Upon closer inspection of our environment and routes we have been
able to identify different types of locations responsible for this. Quiver plots (see
e.g. Fig. 1) reveal that most of these are in regions where paths cross through
virtual tussocks, leading to strong occlusion effects which impair the validity of
image based comparison for all models.
4 Discussion
Here we investigate how a wavelet based representation might be suitable for
view based route navigation with the goal of increasing robustness against lateral displacements. We first observed how RIDFs at single points differ between
pixel space and wavelet space. We found that RIDFs in wavelet space are more
robust and we suggest that this is mainly due to vertical components of high
level details. In order to connect our results with route following we then calculated angular errors along routes with lateral displacements and compared
errors between a selection of wavelet- and pixel-based models. We have shown
that the observed robustness of the wavelet model extends to route following
and outperforms other models while increasing computational efficiency.
Stuerzl et al. [31] investigated how a Fourier coefficient representation of
images influences homing performance. They found that frequency representations of images can be used effectively for snapshot-based visual homing (using
the difference between Fourier components as a proxy for image difference).
These results support the idea of using frequency components for visual navigation. However, a drawback of Fourier analysis is the inability to localise frequencies in an image, which makes it difficult to localise salient image components. An
alternative to Fourier coefficients was introduced in [14], who extracted localised
Haar-like image descriptors at random points in a snapshot and used responses
to these filters to determine familiarity and ultimately derive a homing direction. In contrast, we used filter responses of one level of detail throughout and
applied them to the whole image. Furthermore, our environment, though virtual, was adopted from a real ant site and is notably different from the toy
world used by [31] and the office environment used by [14,31]. Similar to [29],
who observed an increased catchment area when using Zernike Moments, we
observed increased robustness against displacement from the original route when
using wavelet coefficients. The core difference between their work and ours is that
they used summed Zernike Moments, leading to a rotation invariant representation of an image, which wavelets are not. Thus, when compared to the results
reported by [14,29,31] for visual homing, our findings yield additional evidence
that frequency based features are useful and in particular can be applied successfully to route following methods that rely on recovering headings.
In future work, we will optimise the wavelet approach for navigation in the
real world. For instance, we intend to introduce a more elaborate mapping from
S. Meyer et al.
indication that the skyline height without further processing is not a robust
feature for orientation recovery at least in this simulated world.
All models show a noticeable amount of data points outside the 1.5 IQR
interval. Upon closer inspection of our environment and routes we have been
able to identify different types of locations responsible for this. Quiver plots (see
e.g. Fig. 1) reveal that most of these are in regions where paths cross through
virtual tussocks, leading to strong occlusion effects which impair the validity of
image based comparison for all models.
4 Discussion
Here we investigate how a wavelet based representation might be suitable for
view based route navigation with the goal of increasing robustness against lateral displacements. We first observed how RIDFs at single points differ between
pixel space and wavelet space. We found that RIDFs in wavelet space are more
robust and we suggest that this is mainly due to vertical components of high
level details. In order to connect our results with route following we then calculated angular errors along routes with lateral displacements and compared
errors between a selection of wavelet- and pixel-based models. We have shown
that the observed robustness of the wavelet model extends to route following
and outperforms other models while increasing computational efficiency.
Stuerzl et al. [31] investigated how a Fourier coefficient representation of
images influences homing performance. They found that frequency representations of images can be used effectively for snapshot-based visual homing (using
the difference between Fourier components as a proxy for image difference).
These results support the idea of using frequency components for visual navigation. However, a drawback of Fourier analysis is the inability to localise frequencies in an image, which makes it difficult to localise salient image components. An
alternative to Fourier coefficients was introduced in [14], who extracted localised
Haar-like image descriptors at random points in a snapshot and used responses
to these filters to determine familiarity and ultimately derive a homing direction. In contrast, we used filter responses of one level of detail throughout and
applied them to the whole image. Furthermore, our environment, though virtual, was adopted from a real ant site and is notably different from the toy
world used by [31] and the office environment used by [14,31]. Similar to [29],
who observed an increased catchment area when using Zernike Moments, we
observed increased robustness against displacement from the original route when
using wavelet coefficients. The core difference between their work and ours is that
they used summed Zernike Moments, leading to a rotation invariant representation of an image, which wavelets are not. Thus, when compared to the results
reported by [14,29,31] for visual homing, our findings yield additional evidence
that frequency based features are useful and in particular can be applied successfully to route following methods that rely on recovering headings.
In future work, we will optimise the wavelet approach for navigation in the
real world. For instance, we intend to introduce a more elaborate mapping from