Detection of license plates
Automatic number plate detection for ARM platforms

The automatic number plate detection and recognition software (ANPR) is often required in case of intelligent parking systems equipped with cameras for reason of higher security, detection of low behavior, and detailed usage statistics. The need for lowest price of ANPR system leads to centralized solution which allows share both the software and the hardware resources. A disadvantage of such architecture is higher demand on connectivity because of transmission of huge amount of image data produced by each sensing camera separately. This disadvantage can be eliminated by reducing the stream only to images where the number plate is present. To achieve this, we developed and implemented highly optimized C++ shared library for a low-cost camera control computers based on ARM Cortex A7 processor (version with NEON support).


The developed library libANP is designed for real-time application. Input images can be grayscale or color. The ordering of the images has to be preserved and correspond with reality. The output flag about the number plate occurrence is returned for each processed image separately. The library has been evaluated on a dataset composed by 47 937 frames from real traffic scenario. The dataset contains 2 427 positive frames (where the number plate is present), and 45 510 negative frames. The result of evaluation process is summarized in the illustration below:


The total achieved reduction is 90.91% of the original data flow. The probability that the image will be correctly classified is 0.959 (i.e. accuracy of the detector). The development library classified all positive images from dataset properly. The false alarm rate is 0.042. In context of expected usage, this characteristic is desirable. Average computational time on the target platform, ARM Cortex A7, is about 15 ms per image with 752×480 pixels. The memory requirements are minimalistic. The library is also available as module for Python.

Automatic number plate detection
Real-time processing, up to 40 FPS
Optimized for ARM Cortex A7
Available as module for Python
For grayscale and color images

In case of interest, please feel free to contact us at