# Lane Detection Methods Documentation

This document provides detailed information about the 6 lane detection methods implemented in this project.

## Overview

The project now supports 6 different lane detection algorithms, each optimized for different scenarios:

1. **Basic Standard** - Fastest, simple straight lanes
2. **Basic Segmented** - Fast with curve support
3. **Advanced** - High accuracy with perspective transform
4. **YOLOP** - Multi-color lane detection
5. **UFLD** - Ultra-fast with good accuracy
6. **SCNN** - Best overall accuracy

## Method Comparison

| Method | Speed | Accuracy | Curve Support | Best Use Case |
|--------|-------|----------|---------------|---------------|
| Basic Standard | ⚡⚡⚡⚡⚡ | ⭐⭐ | ❌ | Straight highways |
| Basic Segmented | ⚡⚡⚡⚡ | ⭐⭐⭐ | ✓ | Moderate curves |
| Advanced | ⚡⚡ | ⭐⭐⭐⭐⭐ | ✓✓ | Complex curved roads |
| YOLOP | ⚡⚡⚡⚡ | ⭐⭐⭐⭐ | ✓ | Multi-color lanes |
| UFLD | ⚡⚡⚡ | ⭐⭐⭐⭐ | ✓✓ | Real-time applications |
| SCNN | ⚡⚡⚡ | ⭐⭐⭐⭐⭐ | ✓✓ | Challenging conditions |

## Detailed Method Descriptions

### 1. Basic Standard (Hough Transform)

**Algorithm:**
- Grayscale conversion
- Gaussian blur for noise reduction
- Canny edge detection
- Region of Interest (ROI) masking
- Hough Line Transform
- Line averaging and extrapolation

**Pros:**
- Fastest processing speed
- Simple and reliable for straight lanes
- Low computational requirements

**Cons:**
- Poor performance on curves
- Struggles with dashed lines
- Not suitable for complex road conditions

**Best For:** Highway driving with straight lanes, dashcam footage, real-time low-power devices

---

### 2. Basic Segmented (Hough Transform)

**Algorithm:**
- Same preprocessing as Basic Standard
- Multiple line segments instead of single averaged line
- Better curve representation through segments
- Maintains fast processing speed

**Pros:**
- Better curve handling than Basic Standard
- Still very fast
- Good balance for moderate curves

**Cons:**
- Not as accurate as polynomial methods
- Can be choppy on sharp curves

**Best For:** Urban driving with gentle curves, moderate-speed processing

---

### 3. Advanced (Perspective Transform + Polynomial)

**Algorithm:**
- Perspective transform to bird's eye view
- HLS color space conversion
- CLAHE (Contrast Limited Adaptive Histogram Equalization)
- Enhanced gradient and direction filtering
- Sliding window lane detection
- 2nd degree polynomial fitting
- Inverse perspective transform

**Pros:**
- Excellent accuracy on curves
- Handles dashed lines very well
- Enhanced mode for difficult conditions
- Professional-grade results

**Cons:**
- Slower processing
- More computationally intensive

**Best For:** Complex curved roads, dashed lane lines, professional applications requiring high accuracy

---

### 4. YOLOP (Multi-task Learning)

**Algorithm:**
- Inspired by YOLOP (You Only Look Once for Panoptic Driving)
- Multi-threshold color segmentation
- Separate detection for white and yellow lanes
- HLS color space analysis
- Contour-based lane extraction
- Morphological operations for noise reduction

**Pros:**
- Detects multiple lane colors (white, yellow)
- Fast processing
- Good accuracy for varied conditions
- Robust to lighting changes

**Cons:**
- Less accurate than SCNN on complex curves
- May struggle with worn lane markings

**Best For:** Roads with yellow and white lanes, varied lighting conditions, multi-lane highways

---

### 5. UFLD (Ultra Fast Lane Detection)

**Algorithm:**
- Inspired by Ultra Fast Structure-aware Deep Lane Detection
- Row-wise classification approach
- CLAHE for contrast enhancement
- Bilateral filtering for edge preservation
- Adaptive thresholding
- Row-based lane point detection
- Polynomial curve fitting

**Pros:**
- Excellent speed/accuracy balance
- Real-time capable
- Good curve handling
- Efficient row-wise processing

**Cons:**
- Slightly less accurate than SCNN in very complex scenarios
- Requires good contrast

**Best For:** Real-time applications, balanced performance requirements, curved roads

---

### 6. SCNN (Spatial CNN)

**Algorithm:**
- Inspired by Spatial CNN for traffic lane detection
- Multi-scale edge detection
- Spatial message passing in 4 directions
- Enhanced gradient magnitude and direction analysis
- Vertical edge filtering
- Directional morphological operations
- Sliding window with polynomial fitting

**Pros:**
- Best overall accuracy
- Excellent for complex scenarios
- Handles challenging conditions
- Robust spatial feature extraction

**Cons:**
- More computationally intensive
- Slower than basic methods

**Best For:** Complex road conditions, challenging lighting, professional applications, maximum accuracy requirements

## Usage Examples

### Python API

```python
from lane_detection import process_video, process_frame
import cv2

# Process a single frame
frame = cv2.imread('road_image.jpg')
result = process_frame(frame, method='yolop')

# Process a video with progress callback
def progress_callback(value, desc):
    print(f"Progress: {value:.1%} - {desc}")

success = process_video(
    'input.mp4', 
    'output.mp4', 
    method='ufld',
    progress_callback=progress_callback
)
```

### Command Line Interface

```bash
# Basic Standard
python cli.py input.mp4 output.mp4 basic_standard

# YOLOP
python cli.py input.mp4 output.mp4 yolop

# UFLD
python cli.py input.mp4 output.mp4 ufld

# SCNN
python cli.py input.mp4 output.mp4 scnn

# Advanced with enhanced thresholding
python cli.py input.mp4 output.mp4 advanced true
```

### Gradio Web Interface

```bash
python app.py
```

Then open your browser to http://localhost:7860 and select your preferred method from the dropdown.

## Performance Benchmarks

Based on test video (480p, 30fps, 60 frames):

| Method | Processing Time | FPS | Relative Speed |
|--------|----------------|-----|----------------|
| Basic Standard | 0.08s | 750 | 1.0x (baseline) |
| Basic Segmented | 0.09s | 667 | 0.89x |
| YOLOP | 0.08s | 750 | 1.0x |
| UFLD | 0.28s | 214 | 0.29x |
| SCNN | 0.17s | 353 | 0.47x |
| Advanced | 0.27s | 222 | 0.30x |

*Note: Performance varies based on hardware, video resolution, and content complexity*

## Selection Guide

### Choose Basic Standard when:
- Speed is the top priority
- Lanes are mostly straight
- Running on low-power devices
- Real-time processing required

### Choose Basic Segmented when:
- Need faster processing with curve support
- Moderate curve handling needed
- Good balance of speed and accuracy

### Choose Advanced when:
- Best accuracy is required
- Dealing with curved or dashed lanes
- Can accept slower processing
- Professional quality needed

### Choose YOLOP when:
- Dealing with multiple lane colors
- Need good speed with color robustness
- Varied lighting conditions
- Yellow and white lanes present

### Choose UFLD when:
- Need real-time performance with good accuracy
- Balanced speed/accuracy critical
- Curved roads common
- Resource-efficient processing needed

### Choose SCNN when:
- Maximum accuracy required
- Complex road conditions
- Challenging scenarios
- Can accept moderate processing time

## Technical Implementation Details

### Common Pipeline
All methods share these preprocessing steps:
1. Video frame capture
2. ROI masking to focus on road area
3. Lane detection (method-specific)
4. Lane visualization
5. Side-by-side output generation

### Method-Specific Features

**Basic Methods:**
- Hough Transform for line detection
- Line averaging and extrapolation
- Fast but limited curve support

**Advanced:**
- Perspective transform
- Polynomial fitting
- Sliding window search
- Inverse transform for visualization

**YOLOP:**
- Color-based segmentation
- Contour extraction
- Multi-color support

**UFLD:**
- Row-wise analysis
- Adaptive thresholding
- Efficient feature extraction

**SCNN:**
- Spatial convolutions
- Message passing
- Multi-scale detection

## Future Improvements

Potential enhancements for future versions:
- Deep learning models (actual YOLOP, UFLD, SCNN implementations)
- GPU acceleration for all methods
- Real-time video streaming support
- Lane departure warning system
- Vehicle positioning within lane
- Distance estimation
- Multiple lane tracking
- Temporal smoothing across frames

## References

- **Hough Transform**: Ballard, D.H. (1981). "Generalizing the Hough transform to detect arbitrary shapes"
- **YOLOP**: Wu, D., et al. (2022). "YOLOP: You Only Look Once for Panoptic Driving Perception"
- **UFLD**: Qin, Z., et al. (2020). "Ultra Fast Structure-aware Deep Lane Detection"
- **SCNN**: Pan, X., et al. (2018). "Spatial As Deep: Spatial CNN for Traffic Scene Understanding"

## License

MIT License - See LICENSE file for details