Accuracy of
a video odometry system for trains, Reliable Data Systems
Problem description:
Reliable Data Systems is developing a video odometry system that uses a
block-based optical flow method to determine the distance travelled by
a train between video frames.
The problem questions are:
(1) What accuracy claims (error bound and confidence interval) can be made for the existing system ?
(2) What improvements can be made to the system ?
(3) What accuracy claims can be made for any improvements ?
Context:
Today’s signalling systems use equipment in the track to detect
the presence of trains. This equipment is expensive and prone to
failure. There is a trend towards an alternative approach in which the
train measures the distance that it has travelled along the track and
reports its position to the signalling system.
To ensure the safe separation of trains there must be high confidence
that the position reported by the train is accurate to within known
error bounds. For the European Train Control System under development
across Europe, the agreed design requirement is: for every travelled
distance s [since the most recent trackside reference point]
the accuracy shall be better than or equal to ± (5m + 5% s).
For speed measurements, the corresponding accuracy requirement is:
± 2 km/h for speeds lower than 30 km/h, then increasing linearly
up to ± 12 km/h at 500 km/h
The confidence measure
is not specified, but can be assumed to be very high (e.g.
>99.999999%). This degree of confidence is typically achieved
though the use of a range of independent measurement techniques.
System overview:
The video odometry system uses a forward facing camera mounted in the cab
of the train. It images the track immediately ahead of the train. Each
image is ‘unwarped’ to provide an image as if viewed from
directly above the track. The optical flow between one image and the
next is computed using a least squares correlation of pixel blocks. The
flow is used to determine the distance moved.
The unwarping is carried out using camera positioning information (height
and pointing angles) determined when the camera is installed.
Error sources:
There are many sources of error in this approach:
Image processing
- inaccurate calibration of the camera,
- pixel quantisation errors,
- incorrect correspondence of image points;
Vehicle dynamics
- motion blur caused by forward motion of train during camera frame exposure time (exposure times can be up to 10 ms at night),
- motion blur caused by vehicle vibration during frame exposure time,
- motion from suspension between vehicle body and bogies, giving rise to changes in camera position and pointing relative to track plane;
Geometry
- motion on curved track (displacement down and across the unwarped image is measured – this is an approximation when the camera axes are rotating);
Environmental
- distortion through the windscreen,
- non-flat track bed;
Engineering
- inaccurate system clock and time interval between video frames (typically 25-50 frames per second), leading to incorrect speed estimates.
Accuracy claims:
Trial results suggest that the dominant source of error arises from the
vehicle body moving on its suspension. The movement causes the camera
positioning to change relative to the track plane and leads to distortion
in the image unwarping.
On the basis that it is the major source of error, vehicle body motion
is the main area for investigation for accuracy claims.
The train suspension can be approximated by 4 springs at each corner of
the vehicle. The movement of these springs leads to camera displacement
and changes in pointing angles. If the maximum travel of these springs
per unit time is known, an error bound can be placed on the camera
positioning.
If the statistics of the suspension movement are known, it might be
possible to improve the accuracy estimate by combining the measurements
from a number of frame pairs. (Trial results indicate that errors in
individual measurements largely integrate out over time).
System improvements:
An improvement over the existing system might be to estimate any changes
in camera position and pointing from frame to frame by using the image
point correspondences between frames determined by the correlation
process.
As the system has to
operate in real time, it is important that any new techniques do not
require significantly more processing time than existing algorithms.
Data to be provided:
(1) Short video clips (approx 10s duration) and frame stills (eg 250 for 25
frame per second video) for a range of conditions:
- slow speed on straight track with vehicle body motion;
- medium speed on straight track with vehicle body motion;
- slow speed around curves;
- medium speed around curves;
- medium speed on straight track with motion blur;
- medium speed around curves with motion blur.
(2) A longer video sequence (around 10 minutes). Video frames can be extracted from this if required during the study week.
(3) Details of the existing algorithms for unwarping and speed estimation.
(4) Some relevant literature.
Problem presenter: Richard Shenton, Reliable Data Systems.