Lane Detection using CNN

This project focuses on semantic lane segmentation using a U-Net architecture tailored for Indian road conditions. The model is trained on a custom annotated dataset where each pixel is classified into multiple lane categories. U-Net is chosen for its encoder-decoder design, making it ideal for pixel-level segmentation tasks.

The training pipeline supports:

• Multi-class segmentation
• Extensive metric evaluation (Pixel Accuracy, IoU, Precision, Recall, F1-score, Dice Coefficient)
• Automated plotting and saving of training metrics
• Final model saving and result visualization

Architecture of the CNN model used

Dependencies required:

pip install -r requirements.txt

Libraries used:

1. torch==2.6.0
2. torchvision==0.21.0
3. numpy==1.26.4
4. Pillow==10.2.0
5. pyyaml==6.0.2
6. tqdm==4.66.6
7. matplotlib==3.8.4
8. scikit-learn==1.5.1
9. opencv-python==4.11.0.86
10. pandas==2.2.3

Output

Flow of the project

1. Load the Dataset The dataset is stored in the data/processed/ directory and is already split into two sets:

2. Train and Validation Sets

• train/ contains:

  • images/ — input road images
  • masks/ — corresponding segmentation masks

• val/ contains:

  • images/ — input validation images
  • masks/ — corresponding masks for evaluation

3. Train the Model Execute the training script, which:

• Loads the dataset
• Initializes the U-Net model
• Starts training and logs all metrics/graphs to the graphs/ folder

python train.py