Background
Whilst the Leeds High Occupancy Vehicle (HOV) scheme has been judged a success, the cost of on-going manual enforcement has to be borne by Leeds City Council. Clearly, an automated system that could react to illegal occupancy would be advantageous, particularly if the system could also trigger the capture of the registration plate and additional photographic evidence for any subsequent action.
In response to Leeds’ and other Authorities’ difficulty in detecting occupancy, a DfT/EPSRC funded research project has successfully developed a prototype system that can automatically count the number of people in a car. The project partners were Golden River Traffic Ltd, Laser Optical Engineering Ltd, Leeds City Council, Photonics Consultancy, and the University of Sussex.
Leeds HOV lane on the A647
Summary of Research
When the project started the challenge was seen to be the capturing of images of the occupant(s) of a car in such a way that pattern recognition and other image processing techniques could be used to count the number of faces in the car. The original plan was to use a standard video camera operating in the near infra-red camera without distracting the driver (since any illumination at these wavelengths would be invisible). However, modern car windows are heavily coated with layers that absorb at these wavelengths in order to keep the car cool in the sun. Consequently the recognition work concentrated on visible images whilst an alternative infra red wavelength range was investigated.
By making use of the fact that most car windows have a slight blue hue, the area of interest in the scene (the windscreen or side window) could be identified, and the search for faces limited to that area. The three major techniques for face detection were assessed, and a neural network based algorithm was selected as most suitable. Steps were taken to remove artefacts caused by uneven illumination, and with training the system would pick out faces in a moving car.
Suitable infra-red cameras are extremely expensive. However, an important discovery was made. There is a small gap in the absorption of the heat resistant layers on car windows around 1.5 microns. In this region, human skin is very absorbing, whereas hair and clothing reflect the light quite strongly. Thus a human face shows up black (regardless of ethnic origin) whilst the surroundings are grey and white. Dummies (and dogs) can quite easily be distinguished from the real thing. This meant that much simpler processing of the image could be envisaged, so the cost of the camera could be somewhat offset.
The infra red camera works well in bright sunlight, but on a dull day or at night there is not enough natural light at these wavelengths. The only artificial light sources at present are expensive laser diodes. This will not be a problem in the long term as laser costs continue to decline, but meant that this technology alone could not be used for the project demonstrator. However, by combining infra-red and visible images from two cameras the face in a scene can be picked out as a darker blob. Again the camera costs are to some extent offset by the simple processing needed to pull out the faces.
Real time images at different wavelengths of light are collected from the scene (in this case a busy dual carriageway on the outskirts of Leeds) and processed to enhance the contrast between the skin of the face and its surroundings. Software then counts the number of absorbing areas in the scene. In the case of enforcing minimum occupancy a size filter can be applied to ensure that a hand held up will not be counted as the face of a non-existent passenger. If minimum occupancy is not detected a wide angle context photo can be triggered for evidence purposes.
Only one occupant
Output
The project culminated in the demonstration of a working prototype on the High Occupancy Vehicle lane on the A647 into Leeds. Demonstration of an improved version to an invited audience (including representatives from a number of other authorities who are considering installing HOV lanes) is planned for September 2004.
Although the project has been aimed at automatic enforcement of high occupancy vehicle lanes, clearly the technology could also be used at border crossings, or to monitor cars going in and out of high security areas and shopping centre car parks.
Hardware: two cameras and a single lens
For further technical details:
Dr John Brocklehurst, Photonics Consultancy
jrb@photonicsconsultancy.co.uk, +44 1462 743680
For commercial enquiries:
Contact John Brocklehurst as above, who will put you in touch with the relevant organisation.