【车道线检测】Multi-Class Lane Semantic Segmentation using Efficient Convolutional Networks

paper：Multi-Class Lane Semantic Segmentation using Efficient Convolutional Networks

---

Abstract

车道线检测算法流程：lane feature extraction，feature segment grouping, and lane model fitting.

这篇论文提出了两个方法：Feature SizeSelection (FSS) and Degressive Dilation Block (DD Block)，来对现有的语义分割网络进行修正，将现有的语义分割算法拓展到lane markings预测中。

baseline architecture为EDANet

EDANet

Efficient Dense Modules of Asymmetric Convolution for Real-Time Semantic Segmentation

Module

A. Feature Size Selection

The downsampled feature maps tend to lose precise boundary information, and sometimes the thin lanes would vanish.

Therefore, we conclude that placing some convolutional layers at the early stage is able to extract more useful features for small object segmentation since more spatial information is retained

B. Degressive Dilation Block

Dliation convolution may fail to gather the essential local features because of the sparsity of convolutional kernels.

This problem would cause inaccurate recognition of small and thin objects.

Effective Use of Dilated Convolutions for Segmenting Small Object Instances 中提出了一个LFE(Local Feature Extraction)模块，LFE使用degressive dilation rates的空洞卷积来解决上述问题。

Result