View Based Navigation Exploiting Temporal Information
207
Pre-processing: All panoramic images are rendered at an original resolution of
[720×150×3] in RGB. We typically transform the image to a [360×75] greyscale
image using openCV’s resize function using bilinear interpolation. As it has been
shown that resolution of the images can influence homing accuracy [17,27] and
we know that ants have low resolution vision >1 degree per pixel and use light
intensity rather than colour to navigate [9] we transform the image to greyscale
and resize it to what, for an ant, would be high-resolution [360 × 75] (1 degree
per pixel), medium resolution [180 × 50] (2 degree per pixel) and low resolution
[90 × 25] (4 degree per pixel). For pre-processing we compared edge detection
and blurring. For blurring, we use a 5 × 5 Gaussian kernel (low pass filter)
with mean 0 and standard deviation 1 in order to remove high-frequency noise
and smooth out the image. As the kernel is always the same number of pixels,
when we resize the image we effectively scale the amount of smoothing. Fig. 1 c)
shows a panoramic image from AntWorld after resizing and blurring. There is
some experimental evidence [7] that some ants navigate using simple cues on
the horizon as a well defined boundary to derive directional information. As
edge detection can separate sky, tussocks and ground, we also test Canny edge
detection with a lower bound of 180 and upper bound of 200 (pixel values scale
[0–255]) [10].
3 Results
To test the algorithms, we derive headings from a grid of test positions spaced
in 10cm increments around example routes. All variants are tested against 10
different routes and errors are calculated as the absolute difference between the
derived heading and the heading of the nearest snapshot.
Temporal Windows Improve Visual Aliasing Problems: To illustrate the
benefits of our sequence based algorithm, we first compare results on a route
with a loop in it. The agent is given a training route (blue lines in Fig. 1) and we
derive the headings that the agent would follow by each algorithm for all points
within 10 cm of the route. The derived headings are indicated in Fig. 1 by blue
arrows for the ones that are close to correct and by red arrows for large errors
(>40
◦ ). We highlight errors >40
◦ only to demonstrate the behaviour of SMW
against standard PM for a given loop section of a route. In many robotics tasks
this error is not acceptable but here we are most concerned with showcasing the
difference between algorithms and so highlight large deviations from the route
headings.
Errors in orientation when using PM occur because of visual aliasing where
a view will match with a route memory that was stored in a completely different
location [12]. In other words the current view matches with a memory that looks
familiar but is spatially different thus sending the agent in the wrong direction.
There are multiple interlinked reasons for such aliasing. It could be that the view
from two route locations is very similar. To illustrate the problem we look at a
route with a loop.
207
Pre-processing: All panoramic images are rendered at an original resolution of
[720×150×3] in RGB. We typically transform the image to a [360×75] greyscale
image using openCV’s resize function using bilinear interpolation. As it has been
shown that resolution of the images can influence homing accuracy [17,27] and
we know that ants have low resolution vision >1 degree per pixel and use light
intensity rather than colour to navigate [9] we transform the image to greyscale
and resize it to what, for an ant, would be high-resolution [360 × 75] (1 degree
per pixel), medium resolution [180 × 50] (2 degree per pixel) and low resolution
[90 × 25] (4 degree per pixel). For pre-processing we compared edge detection
and blurring. For blurring, we use a 5 × 5 Gaussian kernel (low pass filter)
with mean 0 and standard deviation 1 in order to remove high-frequency noise
and smooth out the image. As the kernel is always the same number of pixels,
when we resize the image we effectively scale the amount of smoothing. Fig. 1 c)
shows a panoramic image from AntWorld after resizing and blurring. There is
some experimental evidence [7] that some ants navigate using simple cues on
the horizon as a well defined boundary to derive directional information. As
edge detection can separate sky, tussocks and ground, we also test Canny edge
detection with a lower bound of 180 and upper bound of 200 (pixel values scale
[0–255]) [10].
3 Results
To test the algorithms, we derive headings from a grid of test positions spaced
in 10cm increments around example routes. All variants are tested against 10
different routes and errors are calculated as the absolute difference between the
derived heading and the heading of the nearest snapshot.
Temporal Windows Improve Visual Aliasing Problems: To illustrate the
benefits of our sequence based algorithm, we first compare results on a route
with a loop in it. The agent is given a training route (blue lines in Fig. 1) and we
derive the headings that the agent would follow by each algorithm for all points
within 10 cm of the route. The derived headings are indicated in Fig. 1 by blue
arrows for the ones that are close to correct and by red arrows for large errors
(>40
◦ ). We highlight errors >40
◦ only to demonstrate the behaviour of SMW
against standard PM for a given loop section of a route. In many robotics tasks
this error is not acceptable but here we are most concerned with showcasing the
difference between algorithms and so highlight large deviations from the route
headings.
Errors in orientation when using PM occur because of visual aliasing where
a view will match with a route memory that was stored in a completely different
location [12]. In other words the current view matches with a memory that looks
familiar but is spatially different thus sending the agent in the wrong direction.
There are multiple interlinked reasons for such aliasing. It could be that the view
from two route locations is very similar. To illustrate the problem we look at a
route with a loop.