A real-time road line detection system optimized for NVIDIA Jetson Nano, designed for autonomous navigation on racing tracks and laboratory environments.

This project implements a lightweight U-Net architecture for semantic segmentation of road lines, specifically optimized for:

• Real-time inference on NVIDIA Jetson Nano (30+ FPS)
• Curved track detection with tight turns
• Yellow and white line recognition
• Varying lighting conditions
• Model size: 15-25 MB for optimal performance

• Lightweight Architecture: Optimized U-Net with attention mechanisms
• Real-time Performance: 30+ FPS on Jetson Nano
• Robust Detection: Handles curved tracks, shadows, and lighting variations

• Encoder: 4 levels with efficient double convolutions
• Attention Mechanism: Lightweight channel attention for better feature focus
• Decoder: Skip connections for precise line reconstruction
• Output: Binary mask (0=background, 1=road lines)

• Input: RGB images (256x256)
• Output: Binary segmentation mask
• Parameters: ~2-5M parameters
• Size: 15-25 MB
• Inference Time: ~33ms on Jetson Nano

project/
├── images/          \# Original track images
├── masks/           \# Binary masks (white=lines, black=background)
└── models/          \# Trained models

# Clone repository

 

# Install dependencies

pip install torch torchvision opencv-python numpy matplotlib tqdm scikit-learn

• Brightness variations: 0.6-1.4x for different lighting
• Contrast adjustment: 0.8-1.3x for line enhancement
• Saturation changes: 0.7-1.3x for yellow/white line distinction
• Rotation: ±12° for curve simulation
• Horizontal flip: 50% probability

• Optimizer: Adam with weight decay (1e-5)
• Learning Rate: 1e-4 with ReduceLROnPlateau
• Loss Function: Binary Cross Entropy
• Early Stopping: Patience of 12 epochs
• Batch Size: 8 (adjustable for GPU memory)

The system processes video frames in real-time and provides:

• Binary mask: Detected road lines
• Control points: Left and right line coordinates
• Steering angle: Calculated direction for navigation
• FPS counter: Real-time performance monitoring
• Visual overlay: Lines highlighted on original image

def extract_control_points(mask):
\# Find contours of detected lines
\# Separate left and right lines
\# Calculate steering angle for MPC/PID
return left_points, right_points, steering_angle

• TensorRT Integration: Convert to TensorRT for maximum speed
• FP16 Precision: Reduce memory usage and increase speed
• Batch Size 1: Optimized for real-time inference
• Memory Management: Efficient GPU memory usage

Input (3, 256, 256)
↓
Encoder: 32→64→128→256→512
↓
Bottleneck: 512 channels
↓
Decoder: 512→256→128→64→32
↓
Output (1, 256, 256)

• Lightweight Attention: Reduces parameters while maintaining accuracy
• Efficient Convolutions: Optimized for mobile deployment
• Smart Augmentation: Track-specific data enhancement
• Real-time Pipeline: End-to-end optimized for speed

• Autonomous Racing: High-speed track navigation
• Laboratory Testing: Controlled environment validation
• Educational Projects: Computer vision and robotics learning
• Research Platform: Algorithm development and testing

🏁 Ready for autonomous navigation on your racing track!