Traffic sign recognition in actual road environment - self-training neural

network target point detection

Video tutorial 42 - training neural network target detection Traffic signs in actual road environment: https://singtown.com/learn/50918/

Code, model, and dataset are all at: https://github.com/SingTown/Traffic-Sign-FOMO

Chinese steps: https://github.com/SingTown/Traffic-Sign-FOMO/blob/main/README-CN.md

Two models were trained in total

The first model: the resolution during training was 128*128, using MobilNetV2 0.35 transfer learning, S1 Score 92%, running frame rate of about 9 frames, and can only run on OpenMV4 Plus. (OpenMV4 does not have enough memory)

The second model: 96*96 resolution during training, using MobilNetV2 0.35 transfer learning, S1 Score 91%, running frame rate of about 12 frames, peak RAM occupancy 244KB, can be run on OpenMV4 and OpenMV4 Plus.

OpenMV partner EdgeImpulse online training website: https://www.edgeimpulse.com/

Note:

1. There are two models in total. If it is OpenMV4, you can only use this small model trained with 96*96 resolution. Plus can use both models.
2. The data set contains five folders, including No Parking x, No Horn n, Speed Limit 80, No Entry-, and Test Pictures. In the actual model and tutorial, we only used No Parking x, No Horn n, Speed Limit 80, and Test Pictures. There is no labeled and trained No Entry- data set. You can choose whether to label and train No Entry-.
3. I only marked the three targets of No Parking x, No Honking n, and Speed Limit 80. I did not mark and train other signs in the dataset (such as Speed Limit 60, Speed Limit 30, No Entry, etc.). Signs that are too similar, such as Speed Limit 80 and Speed Limit 60, may be misidentified. You can also add annotations and train new targets, such as Speed Limit 60, etc.
4. Acknowledgements - Source of China's Actual Road Environment Traffic Sign Dataset (Institute of Automation, Chinese Academy of Sciences): http://www.nlpr.ia.ac.cn/pal/trafficdata/detection.html