License Plate Recognition: Operational Details
BriefCam’s license plate recognition:
Works in a wide range of LPR scenarios and surveillance scenarios
Can read 1 and 2-line plates
Is trained for 4-8 characters per line
Recognizes numbers, Latin characters, and Saudi Arabian Arabic characters.
When there are non-Latin characters in the plate that the user inputs (except for Saudi Arabian Arabic characters), the algorithm ignores them and searches only for the Latin characters (and numbers). If there are non-Latin characters on the plate, the LPR algorithm either ignores the characters or thinks the character is Latin and finds a Latin letter that is similar to the non-Latin character. Since most license plates also include numbers, license plates with non-Latin characters can be searched for using the detected numbers while ignoring the non-Latin characters.
Plate Capture Angles
The BriefCam LPR algorithm has been trained to work at varying angles of capture. However, best results are achieved at small angles when the plate is almost “head-on”.
It is recommended to:
Keep the plate’s yaw and pitch relative to the camera under 40 degrees (larger angles might result in a drop in accuracy).
Keep image rotation to a minimum.
Accuracy
When operating in surveillance “in the wild” scenarios – achieving perfect accuracy is not feasible because of the many challenges as detailed in the BriefCam’s LPR Main Limitations section. Therefore, it cannot be expected that LPR readings in these scenarios will be 100% accurate.
For this reason, we allow searching for plate texts that are similar to a specific query (see the Fuzzy Search and Matching section).
Video Resolution
In general, license plate characters within the plate should be more than 10 pixels in height. The contrast and sharpness should be good. (It is best if the plate is readable to the naked eye in the video.)
Performance
Detection Rate
The BriefCam engine analyzes license plates only in detection frames – which occur at a rate of up to 3 FPS.
Therefore, vehicles need to appear in the scene for at least one third of a second for them to be detected and their license plate analyzed.
Detection Throughput
The LPR engine is limited to extracting a maximum number of 10 plates at once (per frame). Very high activity scenes that have consistently more than five vehicles per frame that need to have their plate detected might require a parameter change (and perhaps additional hardware to take on the additional processing load.