基于渐进式决策融合的快速车道线检测方法

Progressive Decision Fusion Method for Fast Lane Detection

关键词:  车道线检测  多尺度特征挖掘 金字塔特征框架  渐进式决策融合

Key words:  lane detection  multi-scale feature exploration  feature pyramid  framework progressive decision fusion

摘要:

作为自动驾驶领域的一个重要任务，车道线检测旨在通过给定图像找出其中存在的车道线，并对其进行分割。针对当前车道线检测方法在检测尺度较小的车道线时存在易导致车道线部分缺失以及车道线漏检的问题，提出了一种基于渐进式决策融合的快速车道线检测方法。文章首先验证了不同尺度特征在车道线检测任务中的作用，在结构感知阶段引入多尺度特征金字塔模块，挖掘多尺度信息的互补特性，为了提升算法对多种信息的自适应选择效率；其次在车道线检测阶段引入渐进式决策融合模块，在不增加计算量的情况下为了提高算法检测性能。在Tusimple和CULane数据集上的实验结果表明：方法在不显著增加计算量的条件下，相比于基准算法实现了1.2%和6.1%的性能提升，相比于主流算法也有不同程度的性能提升，为车道线实时准确检测提供了可行方案。

Abstract:

As an important task in the field of autonomous driving, lane detection aims to identify and segment the lane lines in a given image. Addressing the issue of partial lane line loss and lane line omission that often occur when detecting small-scale lane lines in current lane detection methods, this paper proposes a fast lane detection method based on progressive decision fusion. The paper first validates the role of different scale features in lane detection tasks and introduces a multi-scale feature pyramid module in the structural perception stage to explore the complementary properties of multi-scale information, thereby enhancing the algorithm's adaptive selection efficiency for multiple types of information. Secondly, a progressive decision fusion module is introduced in the lane detection stage to improve the detection performance of the algorithm without increasing the computational load. Experimental results on the Tusimple and CULane datasets demonstrate that under the condition of not significantly increasing the computational load, the proposed method achieves a performance improvement of 1.2% and 6.1% compared to the baseline algorithm, as well as varying degrees of performance improvement compared to mainstream algorithms, providing a feasible solution for real-time and accurate lane detection.