LLDNet: A Lightweight Lane Detection Approach for Autonomous Cars Using Deep Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Traditional lane detection methods struggle with real-world road variations.
• Deep Learning (DL) models, particularly Convolutional Neural Networks (CNNs), show promise for pixel-level lane segmentation.
• Existing CNN models often lack robustness in adverse conditions and are computationally expensive for embedded systems.

Purpose of the Study:

• To develop a lightweight and robust CNN model for accurate lane detection.
• To address the limitations of existing methods in handling defected roads and adverse weather.
• To create a lane detection system suitable for practical implementation in autonomous vehicles.

Main Methods:

• Introduced LLDNet, a novel, lightweight CNN model with an encoder-decoder architecture.
• Integrated channel and spatial attention modules to refine feature maps.
• Trained the model on a hybrid dataset combining two separate datasets.
• Evaluated performance against state-of-the-art encoder-decoder architectures.

Main Results:

• LLDNet demonstrated superior performance in terms of dice coefficient and IoU compared to existing methods.
• The model achieved outstanding results with a significantly lower parameter count.
• Extensive experiments confirmed accurate lane detection on structured and defected roads, including adverse weather conditions.

Conclusions:

• LLDNet is a highly accurate and efficient lane detection model.
• The model's lightweight design and robustness make it ideal for embedded systems in autonomous driving.
• LLDNet is ready for practical implementation in real-world autonomous vehicle scenarios.