OTVLD-Net: An Omni-Dimensional Dynamic Convolution-Transformer Network for Lane Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Driving Systems

Background:

• Deep learning has advanced lane detection, but current models struggle with challenging scenarios due to limited consideration of unique lane features.
• Existing methods face difficulties and limitations in complex lane topologies and extreme road conditions.

Purpose of the Study:

• To propose a novel lane detection network, OTVLD-Net, that improves adaptability in extreme road conditions and handles complex lane topologies.
• To enhance the extraction of contextual features and aggregate lane symmetry for more robust lane detection.

Main Methods:

• Developed OTVLD-Net using full-dimensional convolutional Transformer, incorporating ODVT-Net with dynamic convolution, feature flip fusion, and non-local network layers.
• Integrated a Transformer-based feature weight generation mechanism, cross-attention, and a vanishing point detection module.
• Employed a joint weighted loss function for coordinated training to boost generalization.

Main Results:

• OTVLD-Net achieved advanced detection performance on OpenLane and CurveLanes datasets, with a 6.4% higher F1 score on OpenLane compared to the second-ranked model.
• Demonstrated an 8.9% average performance improvement in challenging scenarios.
• Achieved real-time performance with 103FPS and 14.2 GFlops using ResNet-18.

Conclusions:

• OTVLD-Net significantly enhances lane detection accuracy and adaptability, particularly in challenging road conditions.
• The model offers a strong balance between high performance and real-time processing capabilities for autonomous driving.
• The proposed methods effectively aggregate global and local features, improving lane detection robustness.