在传统车道线检测算法中，手工设计的特征是算法的关键之一。为了评估不同特征的特点和差异，《Evaluation of Road Marking Feature Extraction》这篇文章中介绍了部分较为基础的图像特征，例如从像素强度或各种滤波响应进行阈值分割得到的像素点，并基于简单的道路标记检测方法进行了对比分析。
framework
graph LR subgraph Evaluation 1[extracting marking features] end 1 --> 2[estimating geometrical model] 2 --> 3[tracking parameters]
Feature extraction Methods: thresholding / gradient analysis / convolution
1. Intro
Six representative road feature extractors and two variants based on thresholding, gradient analysis and convolution were selected. A systematical evaluation of the extraction step is proposed based on manually labeled ground truth.
2. Method Review
Extracting features from a single image captured by a camera in front of a vehicle.
graph LR 1[extractors] --> 2[geometric] 2[geometric] --> 3[segments] 2[geometric] -.- 4[combined] 1 --> 5[photometric] 5 -.- 4 4 --> 6[pairwise gradients] 4 --> 7[convolution filters] 4 --> 8[ridges] 5 --> 9[light pixels] 5 --> 10[edges]
Positive-Negative gradients
extract a line $y{init} \to y{end}$ where $y{end} > T_G, ~ y{init}<-T_G$, when the line lies within the range $[S_m, S_M]$, and the image intensity between the line is higher than the intensity at the two endpoints.
BTW, “Strong gradient” only use local gradient maxima greater than $T_g$
Steerable filters
exploit the spatial continuity of lane markings
Each filter is obtained as linear combination of three basis filters corresponding to the second derivatives of Gaussian distribution.
Top-Hat filters with variant scales
dedicate to vertical lane markings
Global / local / symmetrical local threshold
3. Evaluation
116 road scene images
4. Review
The ROC curves are very confusing, because the FPR (False Positive Rate) became an indicator for better performance in this paper.