基于双分支特征提取的车道线检测算法研究

Study of Lane Detection Algorithm Based on Dual-branch Feature Extraction

摘要:

针对车道线检测算法结构复杂、参数数量较多等问题,本文将车道线检测视为语义分割任务,构建一种基于双分支特征提取的车道线检测模型(简称DP-RESA)。首先,使用两路并行的语义分支和细节分支提取车道线特征。语义分支使用轻量化的Mobilenetv2提取高级特征；细节分支具有更多通道数,用于提取具有更多空间细节的低级特征。其次,利用高级特征的权重对低级特征进行筛选这一思想,快速且高效地融合了两路分支的特征信息。最后,在Tusimple数据集上的实验表明,与baseline相比,本文的DP-RESA算法的准确率提高至96.58%,参数量降至5.12M,单图推理时间降低11.76ms,能够满足车道线检测任务需部署在资源有限的嵌入式平台的需求。

Abstract:

Aiming at the problems of complex algorithm structure and large number of parameters existing in road lane detection, a lane detection method based on dual-branch feature extraction is proposed(abbreviated as DP-RESA). This model treats road lane detection as a semantic segmentation task. Firstly, extract lane features using two parallel branches: a semantic branch and a detail branch. A semantic branch: uses a lightweight Mobilenetv2 model to extract high-level features. A detail branch: with wide channels and shallow layers to capture low-level details with more spatial details. Furthermore, it quickly and efficiently fuses features by using high-level features as weights to filter essential semantic information embedded within the low-level features. Finally, the experimental results on Tusimple dataset show that compared to the baseline, DP-RESA network lane detection accuracy can reach 96.58%, the number of model parameters is reduced to 5.12MB, and lowers single-image inference time by 11.76ms, making it well-suited for lane detection tasks requiring deployment on resource-constrained embedded platforms.