Jetson Parking System / Edge computer vision
From camera frames to plate numbers.
A compact license-plate recognition pipeline with resilient streaming on NVIDIA Jetson.
Camera streams are processed through worker tasks on NVIDIA Jetson.
Illustration of the license-plate recognition pipeline. Plates are fictional examples.
- NVIDIA Jetson
- YOLO
- Redis
- WebSockets
Results & scope
- Camera inputs
- 4
- CV project result
- License-plate recognition pipeline in the August–December 2022 project.
Method & conditions
Self-reported in my CV. Input resolution and Jetson model are not specified.
- Reported recognition-pipeline throughput
- 18 FPS
- CV project result
- Throughput reported for the four-camera parking system.
Method & conditions
Self-reported in my CV. The source does not establish 18 FPS per camera.
- Runtime RAM footprint
- 1.5 GB
- CV project result
- The CV reports 1.5 GB runtime RAM and an 8 MB model footprint.
Method & conditions
Self-reported in my CV. Runtime RAM and model-artifact size are separate measures. Export precision is not specified.
How I approached the work
Engineering decisions.
Recognize plates in two compact stages
- Problem
- Plate localization and character extraction must run within the edge device’s resource limits.
- Decision
- Used a two-stage YOLO pipeline with ordered character boxes, worker-based processing, and Redis-backed streaming with WebSocket reconnects.
- Result
- A four-camera parking pipeline with compact models and a CV-reported 1.5 GB runtime RAM footprint.
The work.
I developed a parking-system pipeline that recognizes license plates from camera streams on NVIDIA Jetson. The work focused on compact detection models, the ordering of detected characters, and reliable communication between processing and application services.
My contribution
Two-stage recognition
Used YOLO detection in two stages to locate license plates and then recognize their characters.
Practical edge constraints
Streamlined models and used bounding-box sorting to extract ordered plate numbers on limited hardware.
Resilient streaming
Applied a worker pattern and WebSocket auto-reconnect, with Redis-backed data streaming.