LHFFNet: A hybrid feature fusion method for lane detection.

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Driving Systems

Background:

• Lane line detection is crucial for autonomous driving but faces challenges like large-scale variations, complex scenes, and high similarity between adjacent lines.
• Remote lane lines are difficult to recognize due to visual angle changes affecting width, leading to classification errors.

Purpose of the Study:

• To propose an effective lane detection framework that enhances multiple spatial features and distinguishes key features across the entire lane line segment.
• To improve the feature representation of lane line images, particularly for distant lane lines, by enhancing and fusing features at multiple scales.

Main Methods:

• A multi-space attention enhancement module utilizing multi-head self-attention to improve the correlation of multiscale lane features.
• A spatial separable convolutional branch with spatial attention weights to emphasize important lane areas in multiscale feature information.
• Introduction of strip pooling before feature output to refine lane line representation and handle anisotropic contexts.

Main Results:

• Achieved 96.84% accuracy on the TuSimple dataset.
• Reached a 75.9% F1 score on the CULane dataset.
• The proposed hybrid network effectively enhances multiscale features for improved lane detection.

Conclusions:

• The developed hybrid feature fusion network demonstrates superior performance in lane detection tasks.
• The integration of multi-space attention, spatial separable convolutions, and strip pooling addresses key challenges in lane line recognition.