Traffic sign classification

Fork on GitHub Watch our video tutorial

Traffic sign recognition with AI can greatly benefit drivers and city infrastructure managers by improving safety, enhancing the driving experience, and increasing efficiency. By using advanced algorithms, the system can accurately detect and recognize traffic signs in real-time, providing drivers or city supervisors with timely and accurate information about speed limits, traffic signals or road sign mapping.

This can help reduce the risk of accidents caused by human error or distraction, making the roads safer for everyone. Moreover, the integration of this technology into transportation systems and traffic monitoring systems can also improve efficiency and reduce congestion.


The use case presented is based on a German traffic sign dataset to classify road signs. The solution proposes models trained on a public dataset providing very good accuracy while running on a STM32.

These models were trained thanks to the Python scripts provided in STM32 model zoo (read our tutorial on the training to learn more).

Here, after setting up our STM32 Model Zoo project, we downloaded the dataset and placed it in the appropriate structure, under

Then we updated the yaml file, which was used to configure the training, with the following settings:

project_name: traffic_sign

name: traffic
class_names: [‘0001’, ‘0002’, …, ‘0042’]
training_path: datasets/train
test_path: datasets/test

Feel free to adjust those settings according to your use-case.

Finally, we simply ran the “” script.

You will find everything you need in the STM32 model zoo to train and retrain any model with your own data. The models can also be easily deployed on a STM32H747 discovery kit with the Python scripts provided in the repository (read our tutorial on the deployment to learn more).

You can find details of all these steps in our Getting started video.


Vision: Camera module bundle (reference: B-CAMS-OMV)


Dataset: The German Traffic Sign Recognition Benchmark*

Data format: The dataset is composed of more than 50 000 RGB images divided into 43 classes of traffic signs.


The STM32 model zoo contains several models for image classification, so you can choose the one best suited to your project.
Here, we have trained 2 different models for the same use-case, each with its own performance and memory footprints:

Model: MobileNetV2 alpha 0.35

Input size: 128x128x3

Memory footprint of the quantized model:

Total RAM: 272 KB

  • RAM activations: 224 KB
  • RAM runtime: 48 KB 

Total Flash: 563 KB

  • Flash weights: 454 KB
  • Estimated flash code: 109 KB


Float model: 99.87%
Quantized model: 99.57%

Performance on STM32H747 (High-perf) @ 400 MHz 
Inference time: 101 ms
Frame rate: 9.9 fps

Model: FDMobileNet

Input size: 128x128x3

Memory footprint of the quantized model:

Total RAM: 65 KB

  • RAM activations: 51 KB
  • RAM runtime: 14 KB 

Total Flash: 192 KB

  • Flash weights: 137 KB
  • Estimated flash code: 55 KB


Float model: 99.78%
Quantized model: 99.51%

Performance on STM32H747 (High-perf) @ 400 MHz 
Inference time: 20.04 ms
Frame rate: 49.9 fps

Confusion matrix of the quantized MobiNetV2 model

* Stallkamp, J., Schlipsing, M., Salmen, J., & Igel, C. (2012). Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. Neural networks32, 323-332.