Aggregate global features into separable hierarchical lane detection transformer

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicle safety relies heavily on accurate lane detection.
• Real-world driving presents challenges like occlusions, poor weather, and faded lane markings.
• Existing lane detection methods struggle with complex environmental factors.

Purpose of the Study:

• To develop an end-to-end lane detection model using a pure Transformer architecture.
• To improve both the accuracy and detection speed of lane detection systems.
• To address limitations of current models in complex road scenarios.

Main Methods:

• Proposed a separable lane multi-head attention mechanism utilizing window self-attention.
• Implemented an extended and overlapping strategy for enhanced inter-window information interaction.
• Developed a pure Transformer-based architecture for lane detection.

Main Results:

• The separable attention mechanism reduces computational cost and increases detection speed.
• The extended overlapping strategy improves global information acquisition and detection accuracy.
• Experimental results demonstrate superior performance over state-of-the-art methods on four datasets.

Conclusions:

• The proposed Transformer model achieves high effectiveness and efficiency in complex lane detection tasks.
• The novel attention mechanism and overlapping strategy are key to improved performance.
• This approach offers a promising solution for robust autonomous vehicle navigation.