End-to-end Lane Detection through Differentiable Least-Squares Fitting
https://github.com/wvangansbeke/LaneDetection_End2End

本文使用 CNN网络来检测车道线,end-to-end 就是输入图像,输出拟合出的车道线参数,一步到位,不用后续处理什么的。

传统的车道线检测是分步骤进行的,一般分为 feature extraction 和model fitting steps

本文的网络结构
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting
图像经过深度网络提取出车道线分布概率图 weight maps,属于车道线的位置其概率值较大,非车道线位置其概率值较小
x-map 和 y-map 分别表示图像中所有像素的 x 坐标和 y 坐标,归一化之后的。结合 weight maps 可以看做是 车道线的位置 x 坐标和 y 坐标,经过 least-squares layer 处理得到 车道线拟合参数

2.2 Weighted Least-Squares Fitting Module
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting
Weighted least-squares fitting
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting

2.3 Geometric Loss Function

车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting

3 Experiments

3.1 Toy Experiment

直线拟合
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting

车道线检测
车道线检测--End-to-end Lane Detection through Differentiable Least-Squares Fitting

111

相关文章: