Automatic Reading of Car Number Plates with Deep Learning

AI-assisted number plate recognition lowers administrative costs and reduces warehouse occupancy

In Switzerland, vehicle number plates are assigned to the owner and not to the car. When deregistering, the road traffic offices must therefore keep the number plates for one year in case the owner wants to register a vehicle again in the same canton. At the road traffic office of the canton of Zurich alone, up to 80,000 pairs of licence plates therefore had to be kept in an automatic warehouse. Renewing the warehouse would have cost millions. A new concept, which relies on computer-assisted automatic identification of the plates, reduces this investment.

In Switzerland, we have a total of 28 categories of vehicle registration plates, which differ in shape, arrangement of the characters, colour and size," explains Stefan Bättig, Head of Infrastructure Projects and Logistics at the Road Traffic Office of the Canton of Zurich. To deregister, the owner only has to put his licence plates in a box in front of the various locations. Up to 1,600 licence plates arrive at the central storage location in Zurich every day. These have to be registered, checked and either archived or destroyed. A critical factor here is the number of licence plates for cars. These accumulate in such quantities that they are stored fully automatically in an automatic warehouse that is now 25 years old. For storage, the plates must first be fixed in stable plastic frames. The warehouse consists of a total of nine heavy-duty paternosters that extend over two floors down to the basement. For the other vehicle categories, manual storage of the licence plates in separate warehouses is sufficient.

"Due to the constant population growth, this paternoster camp was increasingly approaching its capacity limit," Stefan Bättig recalls. When occupancy reached almost 95 per cent in January 2021, a new solution was urgently needed. A new unit would have cost many millions and also required enormous structural expenditure. Therefore, a concept was developed to reduce warehouse occupancy. An analysis of the reactivation rate of the stored signs showed that more than 60 percent of them were not requested again. Therefore, a concept had to be developed that ensured the safe sorting and destruction of no longer needed plates even before they were stored. The prerequisite for this was a division of the delivery containers into two parts. These have two clearly marked delivery slots. Whereas the drop-off for destruction is free of charge, a fee of CHF 40 is charged for plates to be stored. This made it possible to reduce the occupancy rate of the paternoster warehouse to only about 50 per cent within half a year.

Reducing the workload for employees

"The implementation of the new concept, however, first required a far-reaching change in the previous processes for registering incoming signs," says Roberto Accorinti, head of the sign warehouse. Previously, the incoming car plates were always put into storage. In order to be able to recognise them automatically, they first had to be washed to improve their legibility. Bent signs were straightened with the help of a rolling station. Afterwards, signs that belonged together were fastened together in a plastic frame and placed on a conveyor line. From there, each frame was individually guided in front of a camera with the help of a lift mechanism and scanned with the help of character recognition software. This process was very error-prone due to the poor condition of many of the car plates: in about 15 per cent of the cases, the reading had to be corrected manually. Also, clipping the car plates into the frame caused considerable physical strain on the joints and tendons of the fingers, with corresponding adverse consequences for the health of the employees. It would not have made sense to keep this time-consuming procedure for registering the car plates that were destined for destruction from the outset. However, they also had to be registered in any case in order to keep the official keeper database up to date.

Stand-alone camera solution for booking out

"Manual registration of these car plates is not manageable because of the manpower required," adds Roberto Accorinti. Moreover, the error rate would have been far too high. A prerequisite for the successful implementation of the entire project was therefore a dedicated system to register the car plates to be destroyed fully automatically with high reliability and short cycle times. The company Compar in Pfäffikon took on the realisation. The system developed for this purpose consists of a conveyor belt on which the operator - depending on the category - places the car plate or the pair of plates in an oriented manner. The associated computer has a touch screen, a keyboard and other input options. The car plates are conveyed in cycles via a segmented feed section into the camera's recording area, which is shielded from extraneous light, and photographed. For security reasons, two independent software algorithms are used for evaluation: In addition to a classic OCR solution (Optical Character Recognition), the ViDi software from Cognex, which is supported by artificial intelligence (AI), is also used. Compar's own image processing software VisionExpert contains all algorithms, compares the results of both analyses and reacts to differences with a warning signal. The process is stopped and the operator is asked to correct the input. Otherwise, the car plates are ejected. In addition, the programme also detects deviations of the applied signs from the preselected category on the basis of features such as format, colour or arrangement of the car plates and also initiates a rejection in these cases. Correctly identified car plates end up in a large container for transport to a recycling plant.

Reliability rate > 99 percent

"The system was pre-trained at Compar before delivery and therefore already ran almost faultlessly when it was delivered in autumn 2021," says Roberto Accorinti happily. The reliability rate for the recognition of the car plates, which come directly from the drop box of the offices and do not have to be washed or straightened, is above 99 per cent. The operator only has to place the car plates on the belt with the writing facing upwards. Minor deviations in position and angle are automatically corrected by the software. Scanning is so fast that the system cycle only depends on how quickly the operator places new labels. The recognised numbers are stored in an internal database and forwarded to the higher-level IT of the road traffic office for comparison. The few ejected plates can be processed manually.

It was easy for the staff to learn how to use the system. The software created by Compar also proved to be sophisticated, intuitive and easy to use. Compared to the previous system, which is still used for the car plates to be stored, the employees can handle significantly more plates without having to fear health disadvantages.