Logo multitel

Automated inspection of railway traffic recordings

Multitel has been chosen by the French Railway Company to develop, test and validate a prototype for the automated analysis of railway traffic recordings, gathering all the driving information during the journey of a train.

Context of the project

In the railway context, specific train performance parameters are registered thanks to a system of follow-up of driving which is the equivalent of flight recorders in planes. Nowadays, in French railway trains, there are two types of recorders: digital data recorders providing numerical ATESS files, and paper recorders providing rolled up Recording Strips (RS) which are graphic bands of paper containing the following driving information:

  • The time
  • The speed
  • The driving events (DE) such as signal repetition on the locomotive, emergency break, etc.

Each recording strip is associated to a Filling-Card (FC) filled by the train driver for each journey. This card contains all the information related to the successive journeys plotted on the recording strip, for instance date of each journey, driver’s name, origin and destination stations, etc.

(a)      Part of a recorded strip (RS).

(b)     Filling card (FC).

Two procedures are currently used to analyze the 40.000 RS received each month: a partial checking for 100% of them where the operator only pays attention to the main security rules, and a complete randomized checking for only 5% of them, in which each detail of the recording is scrutinized. However, a complete checking for 100% of the RS was recently voted by French law enforcement authorities, which raised the question of an automated and efficient inspection (in time, performance, usability…) of this huge amount of recordings.

In order to automate such inspection process, a machine vision prototype was developed by Multitel, operating the digitization, the extraction and the conversion of the extracted data into the ATESS format, in order to be treated by the same software than the data from the numerical recorders.

Mechanical system

Strip cassette The mechanical part has been thought and designed in order to preserve the thinness and weakness of the recording strip. The concept is largely inspired from the video recorder, with a similar system of cassette and unwinding of the paper. Dedicated cassettes are able to receive any type of RS while the corresponding FC is positioned in the cassette cap, preserving the pair RS/FC during the whole process.
Acquisition system Cassettes are used to digitize both RS and FC using a linear camera, and to manage the RS unwinding.
Overall prototype Upstream, several cassettes are prepared by the operator and stored in a stockroom. The cassette is then automatically introduced in the system and the RS is unwounded under the eye of the linear color camera which scans it, while the FC, located in the cap, is also scanned on both sides. Finally, an output store retrieves the treated cassettes.

Then, we dispose of two different data: the successive images representing the recording strip and the two images of the filling-card. The next step consists in analyzing them and extracting useful information.

Extracting data from recording strips

Concerning the recording strip, the first difficulty is to recognize the type of band, because it conditions the rest of the extraction process, in particular the driving events to detect and the speed limits. The type of band recognition is done by computing a convolution with height templates representing different possible background of RS. According to the RS type, the next step consists in extracting data (covered distance, speed, stops positions…) from the speed/time curves and detecting the corresponding DE. To do so, a segmentation step is applied to extract the interesting time and speed curves from the background with a k-means classifier performed on a new color space. Then, graphs of speed and time are recovered step by step by intersecting the curves with arcs of circle on the segmented image. Finally, the driving events are detected in two main steps: an extraction of all possible events is performed by convoluting the image with three templates representing the basic events’ form, then a post-process is applied so as to filter events with aberrant height, width and density mainly due to writing/acquisition noise. A final step gathers all these information in order to create the output file with the pre-defined format.

The performances of the system have been evaluated in real conditions using the prototype on a large panel of RS which represents approximately 8000 kilometers and 7000 DE. The band type detection gives 100% of good detections. Concerning the DE extraction, results are also quite good, with only 0.28% of non detections and 5 false alarms for 1000km covered by the train. A last evaluation is done on the trains stops which are of primary importance in the main security rules; results give 1.2% of non detections and 3 false alarms for 1000 km covered.

Related publications

C. Machy, X. Desurmont, C. Mancas-Thillou, C. Carincotte and V. Delcourt. Machine vision for automated inspection of railway traffic recordings. January 2009.

V. Delcourt, C. Machy, C. Mancas-Thillou and X. Desurmont. Automatic Reader of Recording Strips. May 2008.