Study group report 2008: accuracy of a video odometry system for trains (Reliable Data Systems)
This is the final report on the problem of the accuracy of a video odometry system for trains, brought to ESGI64 by Reliable Data Systems. Click on the link at the bottom to download the full report as a pdf document.

Report coordinator
Robert Leese (Industrial Mathematics KTN)

Executive summary
Reliable Data Systems is developing a video-based odometry system that enables trains to measure velocities and distances travelled without the need for trackside infrastructure. A camera is fixed in the cab, taking images of the track immediately ahead, at rates in the range 25–50 frames per second. The images in successive frames are ‘unwarped’ to provide a plan view of the track and then matched, to produce an ‘optical flow’ that measures the distance travelled. The Study Group was asked to investigate ways of putting bounds on the accuracy of such a system, and to suggest any improvements that might be made. The work performed in the week followed three strands: (a) an understanding of how deviations from the camera’s calibrated position lead to errors in the train’s calculated position and velocity; (b) development of models for the train suspension, designed to place bounds on these deviations; and (c) the performance of the associated image processing algorithms.

Introduction
In the next generation of railway management systems, such as the European Train Control System (ETCS), there is a trend away from the use of trackside infrastructure to detect the positions and speeds of trains. Reliable Data Systems (RDS) is developing a video-based odometry system that enables a train to measure distances travelled, using a forward-facing camera mounted in the cab. Such a system can report train positions via a radio data link in real time to the signalling control centre, which in turn can provide information to the train on braking points, etc.

In the short term, the benefits of the video system are likely to be in terms of lower costs and higher accuracy of positioning. In the longer term, there are possibilities for allowing closer separation of trains, leading to higher capacity of the rail network.

Alternative approaches suffer from various shortcomings. Systems that rely on trackside infrastructure are expensive to install and maintain. For systems on the train, there may be limitations on performance in winter conditions, for example due to wheel slip for devices that measure wheel rotations. Inertial methods are expensive. There has been much work done on satellite positioning, but visibility is not consistent (for example in tunnels) and so a secondary sensor system is needed to provide coverage during those periods. The video system being developed by RDS appears to overcome all these problems.

The overview operation of the video-based system is as follows. The camera mounted in the cab images the track immediately ahead of the train, generally at a frame rate in the range 25–50 frames per second. Each image is ‘unwarped’, to provide a plan view as if viewed from directly above the track. The unwarped images from successive frames are matched by looking at pixel blocks, to build up an ‘optical flow’ from one image to the next. This flow provides an estimate of the distance moved between frames.

Early trials have indicated that this technique is robust, and works effectively at a range of train speeds and in a variety of weather and lighting conditions. However, in order for such a system to be adopted by the industry, it is important to have (very) high confidence that the positioning is accurate to within known error bounds. In ETCS the agreed design requirement is that for a distance travelled s (from some reference point) the accuracy in position shall be better than ± (0.05s + 5 metres). For estimates of speed, the accuracy requirement is ± 2 km/h for speeds up to 30 km/h, then increasing linearly up to an accuracy of ± 12 km/h at a speed of 500 km/h. Empirically, the accuracy of the current RDS position estimates is better than ±0.025s, but it is important to provide a rigorous underpinning for such claims.

Challenges for the study group
The three questions addressed to the Study Group were:

1. What accuracy claims can be made for the existing system?
   
2. What improvements might be made to the existing system?
   
3. What accuracy claims can be made for any improvements?

Sources of error
The possible sources of error in the systems are:

1. Imaging deficiencies: these include inaccurate calibration of the camera, errors in the timing of video frames, and motion blur due to finite exposure times. These are not felt to be the major sources of error, and can in any event be addressed through upgrading the camera hardware.
   
2. Vehicle body motion: trial results indicate that this is the dominant source of error, and is where the Study Group focussed its attention. The position and orientation of the camera will in general be affected by movements in the train suspension. In the current system, the unwarping is fixed, following initial calibration. Therefore small changes in position and orientation will mean that the unwarping is slightly in error, and these errors will propagate to the estimates of position.
   
3. Nonplanar track bed: the positioning calculations assume that the track bed is planar. In practice, sleepers and ballast are ‘bumpy’, or there may be uneven snow cover. There may be also be changes in track gradient and occasional sharper raised areas, such as those at level crossings.
   
4. Image processing: errors can result from the incorrect correspondence of image points (for example caused by lack of detail on the imaged surfaces) and pixel quantisation. The effects of vehicle body motion on the image processing algorithm were also looked at in the Study Group. The need for real-time operation is a constraint on the computational complexity of the image processing.
   
5. Cornering: errors due to the additional rotational motion when cornering are a particular case of errors due to more general vehicle body motion.

Download 'RDS-AccuracyOfAVideoOdometrySystemForTrains.pdf' (420 Kb).