Basic Concepts

What Can You Do With degirum-vehicle?

degirum-vehicle enables license plate recognition workflows across images and video:

• Detect and recognize license plates in photos, video files, and live streams

• Track vehicles across video frames with persistent IDs

• Improve accuracy via Bayesian multi-frame text aggregation

• Real-time alerts when license plates are detected

• Batch processing for large image collections

• Automated clip extraction for detected vehicles

All with minimal code and flexible deployment options.

Deployment Options

Where Can It Run?

The inference_host_address parameter controls where your models execute:

Option	Description	Best For
Cloud (@cloud)	Models run on DeGirum's cloud servers	Quick start, experimentation, trying different hardware
Local (@local)	Models run on your machine	Production, lowest latency, full privacy, offline operation
AI Server	Models run on a dedicated server (e.g., server-ip:port)	Centralized GPU/accelerator resources on your network

Cloud is ideal for getting started quickly without hardware setup. Local gives you complete control and privacy. AI Server lets you share powerful hardware across multiple applications. See AI Server Setup Guide for server deployment.

What Hardware Is Supported?

The device_type parameter specifies which accelerator to use. The following hardware platforms are supported (listed in alphabetical order by vendor):

Hardware Platform	Device Types	Description
CPU	TFLITE/CPU	Standard CPU inference (baseline)
Axelera Metis	AXELERA/METIS	Axelera Metis AI accelerator
DeepX	DEEPX/M1A	DeepX M1A accelerator
DeGirum Orca	N2X/ORCA1	DeGirum Orca AI accelerator
Google Coral	TFLITE/EDGETPU	Google Coral Edge TPU
Hailo	HAILORT/HAILO8 HAILORT/HAILO8L	Hailo-8 and Hailo-8L AI accelerators
Intel OpenVINO	OPENVINO/CPU OPENVINO/GPU OPENVINO/NPU	Intel CPUs, integrated GPUs, and NPUs

See DeGirum PySDK Installation for platform requirements and Runtimes & Drivers for hardware-specific setup.

Discovering Available Hardware

Use helper functions to check what's supported and available before configuring:

# See all supported hardware platforms
supported = degirum_vehicle.model_registry.get_hardware()

# Check what's available on cloud or local
available = degirum_vehicle.get_system_hw("@cloud")

# Find compatible hardware you can use right now
compatible = degirum_vehicle.get_compatible_hw("@cloud")

Model Registry

degirum-vehicle includes a curated model registry - a collection of pre-optimized license plate detection and OCR models for various hardware platforms. The registry automatically selects the best model for your chosen hardware, eliminating the need to manually pick models or tune parameters.

Helper functions like get_license_plate_detection_model_spec() and get_license_plate_ocr_model_spec() query this registry to get optimized models for your hardware. For complete control, you can also provide custom models outside the registry.

Core Components

degirum-vehicle provides three main components:

LicensePlateRecognizer - For static images and batch processing

• Process photos or image collections

• Detect and recognize license plates

• One-time batch recognition

LicensePlateTracker - For video streams and real-time monitoring

• Monitor live camera feeds or video files

• Track vehicles across frames with persistent IDs

• Bayesian multi-frame text aggregation for improved accuracy

• Generate real-time alerts and automated clip extraction

LPRClipManager - For managing saved video clips

• List and retrieve video clips from object storage (S3 or local)

• Access clips recorded by LicensePlateTracker alerts

• Download and manage alert recordings

---

Before diving into the component-specific guides, let's understand the underlying recognition pipeline that powers all these tools.

The Recognition Pipeline

Both LicensePlateRecognizer and LicensePlateTracker use the same core pipeline for processing license plates:

┌─────────────┐     ┌──────────────────┐     ┌──────────────┐     ┌──────────────┐
│ Input Image │ ──> │ License Plate    │ ──> │   Crop &     │ ──> │     OCR      │
│             │     │   Detection      │     │   Extract    │     │  Extraction  │
└─────────────┘     └──────────────────┘     └──────────────┘     └──────────────┘

1. License Plate Detection

Locates license plates in the image and provides bounding boxes for each detected plate.

Output: Bounding box coordinates per plate

2. Cropping & Extraction

Extracts each detected license plate region and resizes to standard OCR input size.

Output: Cropped plate image (typically 256×128 pixels)

3. OCR Extraction

Converts the cropped plate image into recognized text using character detection.

Output: Plate text string + confidence score

4. Bayesian Aggregation (Tracker Only)

For video streams, combines OCR results across multiple frames using exponential moving average of character probabilities.

Output: Aggregated text + confidence score per vehicle track

Additional Capabilities

Zone Filtering (optional) - Applied after detection to process only plates within specified polygon regions. See Vehicle Filters Reference.

Tracking (LicensePlateTracker only) - Maintains persistent IDs for vehicles across video frames using visual tracking. Enables features like trajectory tracking, temporal consistency, and multi-frame text fusion.

Alert & Recording (LicensePlateTracker only) - Triggers events when license plates are detected and automatically records video clips to object storage (S3 or local filesystem).

Understanding Results

Results are returned as LPRResult objects containing detected plates with their text and metadata:

result = lpr.predict("car.jpg")

for plate in result.license_plates:
    print(f"Plate: {plate.plate_number} (OCR: {plate.ocr_score:.2f})")

Key properties:

• plate_number - Recognized text (or None if unreadable)

• ocr_score - OCR confidence (0.0-1.0)

• detection_score - Detection confidence (0.0-1.0)

• bbox - Plate bounding box

See LPR Data Classes Reference for complete property list and usage examples.

---

Next Steps

Now that you understand deployment options, hardware support, the core components, and how the recognition pipeline works, you're ready to start using degirum-vehicle.

Choose your path:

• Process images or batches → LP Recognizer Guide

• Monitor video streams → LP Tracker Guide

• Learn about models → LPR Models