HWLane

Lane Detection using HW-Transformer

HWLane: Transformer-based Lane Detection for Autonomous Driving

Research Overview

This project presents HWLane, a novel transformer-based approach for lane detection in autonomous driving. The model leverages self-attention mechanisms across both row and column dimensions to capture complex spatial relationships in road scenes.

By combining horizontal and vertical attention patterns, HWLane achieves superior performance in detecting lane boundaries under various challenging conditions including occlusions, shadows, and varying lighting.

Key Features

  • •HW-Transformer: Dual-axis self-attention mechanism
  • •Robust Detection: Handles occlusions and shadows
  • •Real-time Performance: Optimized for autonomous systems
  • •Improved Accuracy: State-of-the-art results on benchmarks

Datasets Used

TuSimple

Preprocessed dataset with 6,408 training images and 2,550 test images

Resolution

1280 × 720 pixels with high-quality lane annotations

Augmentation

Data augmentation techniques applied for robustness

Model & Results

Performance metrics and architecture comparison

Model Architecture

Baseline: Tiny U-Net

Lightweight encoder-decoder with skip connections

Improved: HW-Transformer

Integrates horizontal and vertical self-attention layers

Enhancements

SWA (Stochastic Weight Averaging) & TTA (Test Time Augmentation)

Architecture Distribution

Performance Metrics Comparison

IoU Score Progression

Best IoU Score

0.5210

U-Net + TTA

Accuracy

93.10%

Improved Model

F1-Score

0.9050

Best Performance

Code & Implementation

Explore the implementation details and code snippets

Full Implementation

Access the complete codebase on GitHub

View on GitHub

Interactive Visualization

See lane detection in action

Highway Scene

Input image

How It Works

  1. 1.Input image is preprocessed and resized to 720×1280
  2. 2.HW-Transformer processes the image with dual-axis attention
  3. 3.Decoder generates lane detection mask
  4. 4.Post-processing refines lane boundaries

Performance Stats

Inference Time~45ms
Detection Accuracy93.1%
Model Size12.5 MB