Skip to content

About

World's fastest ANPR / ALPR implementation for CPUs, GPUs, VPUs and NPUs using Transformers (ONNX, Tensorflow lite, TensorRT, Hailo, OpenVINO, RKNN2...). Multi-Charset (Latin, Korean, Chinese) & Multi-OS (Jetson, Android, Raspberry Pi, Linux, Windows) & Multi-Arch (ARM, x86).

Topics

Resources

Stars

751 stars

Watchers

29 watching

Forks

Repository files navigation





UltimateALPR is a highly optimized ANPR / ALPR (Automatic Number/License Plate Recognition) SDK designed to run offline on embedded devices (cameras, ARM64 mobile devices...) and servers (X86_64). The SDK focuses on speed and accuracy with a particular touch on low memory usage thanks to quantized models (QUINT8).

The SDK supports multiple deeplearning frameworks (Tensorflow Lite, OpenVINO, NVIDIA TensorRT, Hailo SDK, RKNPU2, ONNX Runtime...) to allow targeting various types of GPUs and NPUs.

The latest version (v4.0.0) is a complete re-design and focuses on broad support for different NPUs and GPUs to maximize speed:

  • NVIDIA GPUs (Desktop, Jetson...) via ONNX Runtime (cuda) and TensorRT
  • Intel GPUs, NPUs and FPGAs via OpenVINO
  • Rockchip NPUs (radxa rock 4b, orange pi...) via rknpu2 framework
  • Hailo NPUs
  • Qualcomm, Verisilicon (NXP i.MX 8M Plus, Khadas, STMicroelectronics...) and many others NPUs thanks to Tensorflow Lite external delegates.

With a minimum frame rate as high as 37 fps on Raspberry Pi 5, 91 fps on Galaxy S24, 310 fps on Orange Pi 5 and 472 fps on Jetson Nano Orin, this is the fastest ANPR/ALPR implementation you'll find on the market. Being fast is important but being accurate is crucial. We use state of the art deep learning techniques (multi-head attention, transformers...) to offer unmatched accuracy and precision.

The next video shows the Recognizer sample running on Windows:
Recognizer Running on Windows


In addition to License Plate Recognition (LPR) we support Image Enhancement for Night-Vision (IENV), License Plate Country Identification (LPCI), Vehicle Color Recognition (VCR), Vehicle Make Model Recognition (VMMR), Vehicle Body Style Recognition (VBSR), Vehicle Direction Tracking (VDT) and Vehicle Speed Estimation (VSE).

Don't take our word for it, come check our implementation. No registration, license key or internet connection is required, just clone the code and start coding/testing. Everything runs on the device, no data is leaving your computer.

You can also check our online cloud-based implementation (no registration required) to check out the accuracy and precision before starting to play with the SDK.

Trying samples (C++, C#, Java and Python)

Go to the samples folder and choose your prefered language. You don't need to build the samples by yourself, the binaries folder contains pre-built apps to quickly test the SDK.

Benchmark

This SDK is the fastest you'll find on the market, we provide pre-built binaries to check the speed on your device. Here are some benchmark numbers on most commonly used ARM64 edge devices:

ARM64 edge Device Frame rate
Jetson Nano Orin (TensorRT GPU) 472.89
Dragon Q6A (Qualcomm Hexagon NPU) 317.18
Orange Pi 5 (RKNPU2) 310.21
Raspberry Pi 5+ (Hailo-8L NPU) 143.67
Samsung Galaxy S24 (TfLite CPU) 91.24
Debix A (NXP i.MX 8M Plus NPU) 71.02
Raspberry Pi 5+ (TfLite CPU) 37.41

More benchmark numbers could be found at samples/c++/benchmark.

Technical questions

Please check our discussion group or twitter account

About

World's fastest ANPR / ALPR implementation for CPUs, GPUs, VPUs and NPUs using Transformers (ONNX, Tensorflow lite, TensorRT, Hailo, OpenVINO, RKNN2...). Multi-Charset (Latin, Korean, Chinese) & Multi-OS (Jetson, Android, Raspberry Pi, Linux, Windows) & Multi-Arch (ARM, x86).

Topics

Resources

Stars

751 stars

Watchers

29 watching

Forks

Releases

Packages

Used by

Contributors

Languages