frame processing
This element is part of the full workflow: from input data to review and result.
security and video monitoring
License Plate Recognition on RK3568
An autonomous license plate detection and recognition system for local video processing on energy-efficient RK3568 devices.
data loop production-ready
01 Data
02 Processing
03 Review
04 Report
API integrations and statuses
Reports PDF, XLS, BI
AI validated on data
task
What had to be solved
Recognize Russian license plates without cloud processing: detect the plate in a frame, read the characters and pass the event to an external system.
The real project name, client, site, plate numbers, coordinates, internal URLs and identifiers are not disclosed.
constraints
What we had to account for
- real plate numbers had to be hidden
- limited RK3568 resources
- processing speed requirements
- Russian state license plate format
- different cameras and lighting conditions
Project loop
01 license plate area detection
02 modified LPRNet architecture for the Russian plate format
03 inference optimization for RK3568
04 local event service
05 API for result transfer
implementation
What was built
license plate detection
This element is part of the full workflow: from input data to review and result.
character recognition
This element is part of the full workflow: from input data to review and result.
LPRNet architecture adaptation
This element is part of the full workflow: from input data to review and result.
event logging and transfer
This element is part of the full workflow: from input data to review and result.
result
What the business gets
- up to 2.5 frames per second on RK3568
- up to 97% accuracy on Russian license plates
- recognition accuracy increased from 70% to 97%
- the system works autonomously without the cloud
- the result is ready for integration into an external workflow
reuse
What applies to similar tasks
- license plate recognition
- edge inference
- model optimization for constrained devices
- local processing loop
- event API
stack
Technologies and approaches
OpenCVONNXPythonLPRNetRK3568local processing
next step
Need to solve a similar task?
Describe the data, current workflow, constraints and expected result. We will suggest a safe format for the first validation.