Law enforcement

Law enforcement applications provide the greatest challenge for license plate reading. Three major reasons for this are the extremely low error rate requirements (typically below 0.5%), the need for large images containing many visible details of the environment, and the virtually unbounded position of the license plate in the image. As an additional difficulty, these large images may contain multiple vehicles, in which case the license plates of all these vehicles should be read.

This implies that the entire image needs to be scanned for one or more license plates. Furthermore, because of varying triggering conditions, a wide range of possible character heights (typically between 12 and 32 pixels) has to be dealt with simultaneously.

Intrada ALPR uses efficient and fast algorithms for finding the correct license plate locations and multiple classifier techniques to obtain a reliable, country specific, readout.

As an example some law enforcement images have been added. Original image sizes vary around 3000x2000 pixels. There are often multiple vehicles present, and there is a large amount of environmental detail. Both camera viewing angle and lighting technique (visible light, infrared and infrared flash) are non-fixed.

Free flow tolling

Free flow tolling applications involve fast moving vehicles on multiple lanes and varying weather and lighting conditions. Cameras are typically mounted over the road, at for instance a portal, a gantry or a bridge. External triggering results in a fairly constant vertical location of the license plate in the input images.

As imaging conditions are less controllable than with parking, the typical recognition rate is somewhat lower.

Several tollgate images are shown, which are taken by professional traffic cameras from vendors like JAI PULNiX, using infrared lighting.

Parking

Parking applications typically operate with slow moving, or stopped, vehicles at a barrier or gate. This implies small variations in the license plate location and size within the camera image. Any perspective distortion is fixed and can be compensated automatically by Intrada ALPR. The optimal character height of 15 pixels can be obtained simply by setting the appropriate zooming factor.

Also, lighting conditions at the gate are rather constant and thus independent of weather and time of day. Camera parameters need only be set once to fully exploit these invariant lighting conditions for optimal license plate reading.

Therefore parking applications can deliver constant image quality and invariant image contents, which can both be strongly optimised, resulting in an almost 100% recognition rate by Intrada ALPR.

Some typical parking input images, taken at different gates, are shown below to illustrate the above mentioned conditions. These half-frame images are taken by a standard CCTV camera (Philips LTC 500).

Conclusion

The error-correct performance curves for the three application domains, are shown below. These curves are typical for each domain, but the actual shape may vary depending on the camera setup and imaging conditions.

The user can select an arbitrary operation point on any of these curves by setting the appropriate confidence level of the final recognition result.

The current Intrada ALPR release might provide better results than shown here, since Dacolian is constantly improving the performance curves for all application domains towards the lower right corner of the error-correct graph.