Install requirement - pip install -r requirements.txt

Traffic sign detection and recognition

Our project is focused on recognizing traffic signs using data from the Mapillary Traffic Sign Dataset. Our main focus was on fine-tuning the YOLOv8 model, which tends to produce state-of-the-art results for many object detection tasks in real time. In principle, we employ three different approaches. The first approach involves a one-step process, utilizing a YOLOv8 model for simultaneous traffic sign detection and classification. The second approach employs two separate YOLOv8 models — one for binary detection (sign/no-sign) and another for classification of the pre-detected sign. The third approach involves fine-tuning the Object detection transformer DETR. See the documentation for further details.

Data preprocessing

• download dataset and edit the structure of it according to documentation
• edit PATH variable in data.py
• run the data.py script

Model trainings

• edit path and other constants at the beginning of files
• run
  • model.py to train binary detector and yolo classifier models or yolo end to end detector and classifier model
  • train_cls.py to train simple CNN classifier
  • train_detr.py to train DETR model

Evaluation

• run eval_double_step to run decoupled approach evaluation
• to run evaluation of simultanious approach, load model using model = YOLO('path/to/best.pt') and run model.eval()

Inference with visualisation

• edit paths to image and models in inference.py
• run inference.py

Acknowledgments and Links

Notice: The complete GitHub repository exceeded the size limit of the assignment. Therefore, we are providing you with a link to access the repository hosted at https://github.com/.

• The GitHub repository is accessible here
• The fine-tuned models are available here
• The code for the DETR was adapted from this