Fast and Accurate Adaptive Lane Detection via Curve Modeling

Lane line detection is a core component of the visual navigation system for intelligent vehicles. In ideal driving scenarios, lane lines exhibit the characteristic of left-right symmetry, and the horizontal flipping and subsequent fusion of the lane lines on one side can be used to assist in detecting the lane lines on the other side that are difficult to identify. However, in actual driving scenarios, most lane lines are not completely symmetric, and there is an offset in the lane line features between the feature maps before and after flipping. In order to tackle this challenge, we introduces an adaptive flipping fusion module to bridge the feature gap between the feature maps before and after flipping. In addition, considering that the model is based on the Bézier curve modeling, in order to make the most of the characteristics of the Bézier curve, we design a loss calculation method with a non-uniform sampling point distribution to optimize the model training. We conducted a host of experiments on the CULane dataset which show that our method achieves fast and accurate lane line detection with the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex> score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 5. 9 5 \%}$</tex>, ensuring robust technical support for the practical implementation of ADAS and autonomous vehicle systems.