LP Tracker

What is LicensePlateTracker?

LicensePlateTracker processes video streams with real-time license plate detection, tracking, and recognition. It maintains persistent vehicle identities across frames and uses Bayesian multi-frame text aggregation to improve OCR accuracy.

When to Use LicensePlateTracker

Choose LicensePlateTracker when you need to:

• Process live camera feeds or RTSP streams in real-time
• Track vehicles with persistent IDs across video frames
• Improve OCR accuracy via Bayesian multi-frame text fusion
• Generate alerts and notifications for detected plates
• Save video clips with detected vehicles

For simple batch image processing without temporal tracking, use LicensePlateRecognizer instead.

Core Concepts

Configuration

All settings are controlled through LicensePlateTrackerConfig:

import degirum_vehicle

config = degirum_vehicle.LicensePlateTrackerConfig(
    license_plate_detection_model_spec=detection_spec,  # Detection model
    license_plate_ocr_model_spec=ocr_spec,             # OCR model
    video_source="highway.mp4",                        # Video source
    live_stream_mode="LOCAL",                          # Display mode
)

tracker = degirum_vehicle.LicensePlateTracker(config)

See Configuration Guide for all options.

Pipeline Architecture

The tracker uses a multi-stage streaming pipeline:

1. Video Source - Reads from camera, file, or RTSP stream
2. License Plate Detection - Detects plate regions in each frame
3. Vehicle Tracking - Tracks vehicles across frames with unique IDs
4. Cropping - Extracts detected plate regions
5. OCR Recognition - Runs text recognition on cropped plates
6. Bayesian Fusion - Aggregates OCR results across multiple frames per vehicle track
7. Annotation - Overlays tracking IDs and plate text on video
8. Output - Display, save to file, or stream via RTSP

The Bayesian fusion uses exponential moving average of character probabilities to improve accuracy over single-frame recognition.

Key Features

Vehicle Tracking

• Unique track IDs for each vehicle
• Track persistence across frames
• Zone-based filtering (optional)
• Track trail visualization

Multi-Frame Bayesian Text Aggregation

• Combines OCR results across multiple frames
• Exponential moving average of character probabilities
• Improved accuracy vs. single-frame recognition
• Confidence scoring per track

Video Processing

• Live camera feeds (webcam, IP camera)
• Video files (MP4, AVI, etc.)
• RTSP streams
• Real-time display or RTSP streaming output

Event Notifications

• Configurable alert conditions
• Apprise notification support (email, Slack, Discord, etc.)
• Video clip storage with alerts

Methods

LicensePlateTracker provides methods for video processing:

• start_tracking_pipeline() - Start real-time tracking pipeline
• predict_batch() - Process video stream and return results
• find_plates_in_file() - Analyze entire video file
• find_plates_in_clip() - Analyze video from object storage

See Methods Reference for complete API documentation with examples.