The license plate recognition module has been significantly improved. Different neural network models are now used for European, American, and other plates, resulting in a substantial increase in recognition accuracy.
Special attention has been given to east-europe license plates: the last 2–3 digits, which indicate the federal region, use a different font and positioning. A separate optimized neural network is applied for these cases.
The system now automatically detects the plate’s country of origin and selects the most suitable recognition model. The program has been successfully tested on most European license plates. Additionally, new parameters have been added to the ini-file, allowing for flexible fine-tuning of the recognition process.
CPU load has been optimized — an especially important improvement when working with a large number of streams and cameras.
IMPROVED MOTION DETECTION & VIDEO PLAYBACK
The neural network–based motion detector has been thoroughly redesigned. It is now more accurate and more responsive to real changes in the frame, while the number of false alarms has been significantly reduced.
A bug affecting the live view of certain high-resolution IP cameras has been fixed. For such cameras, the buffer size has been increased to ensure more stable operation.
Playback of recorded video is now performed in the new SmartVision Player, which provides a smoother and more comfortable navigation experience through h.265 records.
WHAT AFFECTS LICENSE PLATE RECOGNITION QUALITY
License plate recognition is always a balance between accuracy and processing speed. The result depends on multiple factors:
Camera positioning
Frame rate (FPS)
Lighting conditions
Computer performance
The higher the frame rate, the greater the chance of accurate recognition. For example, at 30 FPS the accuracy is noticeably higher than at 5 FPS. However, increasing the frame rate also requires more processing power.
Even with a high frame rate, incorrect camera placement can lead to situations where, at high vehicle speeds, the detector “sees” only a single frame with the plate — and recognition fails.
To successfully identify a plate, the system must receive several matching recognitions. The required number of matches can be configured in the INI file.
Additionally, algorithms analyze the probability of character-level recognition errors to determine the most likely plate number. Vehicle tracking is also applied to avoid confusion when multiple cars appear in the frame.
When setting up the recognition module, special attention should be given to proper camera placement and the relative speed of vehicles in its field of view. In individual frames, plates may be blurred or distorted — this is why the system processes multiple frames, averages the results, and applies intelligent decision-making algorithms.
The higher the accuracy setting, the better the camera and the more powerful the computer must be for stable operation.
EXAMPLE: DISTANCE TRAVELED PER SECOND
20 km/h — 5.56 m
40 km/h — 11.11 m
60 km/h — 16.67 m
120 km/h — 33.33 m
200 km/h — 55.56 m
Even at 20 km/h, if the camera is placed too close to the license plate, the system may not capture enough frames for confident recognition, even though the plate looks clear to the human eye.
This issue can be resolved in two ways:
Adjusting recognition parameters — reducing the required number of frames to make a decision.
Repositioning the camera — installing it so that the plate remains in the field of view for at least one second.
For fine-tuning the parameters and optimal placement of equipment, assistance from a qualified engineer may be required.