带语义分割的轻量化车道线检测算法

Lightweight Lane Detection Algorithm with Semantic Segmentation

为了解决在计算资源有限的车载嵌入式设备中车道线检测算法存在实时性差、精度不高的问题,提出了一种带语义分割的轻量化车道线检测算法(SegLaneNet).首先通过简化并联的空洞卷积支路,增加跳跃连接结构,提出新的空洞空间金字塔池化模块(ASPP-tiny);接着定义模型的多尺度输入、跳跃连接的浅层特征与深层特征融合、并联不同采样率的空洞卷积特征融合;再有对自编码器中的上采样与下采样卷积进行剪枝操作,提出一种新的轻量化全卷积语义分割算法SegLaneNet应用于车道线检测;最后与Baseline算法相比,本文的SegLaneNet算法在图森(TuSimple)车道线检测挑战数据集上测试的准确率提高了约2%,假正例(FP)减少了3%以上,假负例(FN)减少了约2%.在GPU服务器上测试运行速度达165帧/秒(FPS),同时在嵌入式设备中运算速度达到16帧/秒(FPS).测试结果表明带语义分割的轻量化车道线检测算法能够满足车载嵌入式设备实时、准确的车道线检测工作.

In order to solve the problems of poor real-time and low accuracy performance of lane detection algorithms in vehiclemounted embedded devices with limited computing resources,a lightweight lane detection algorithm(SegLaneNet)with semantic segmentation is proposed.First,by simplifying the parallel hollow convolution branches and adding jump connection structures,a new atrous spatial pyramid pooling module(ASPP-tiny)is proposed.And then pruning the up-sampling and down-ampling convolution in the autoencoder,a new lightweight fully convolutional semantic segmentation algorithm SegLaneNet is applied to lane line detection.Finally,compared with Baseline algorithm,the accuracy of the SegLaneNet algorithm on the TuSimpIe lane line detection challenge dataset has been improved by about 2%Jthe false positive(FP)has been reduced by more than 3%Jthe false negative(FN)has been reduced by about 2%,and the running speed on the GPU server has reached 165 frames per second(FPS),while The computing speed in embedded devices reaches 16 frames per second(FPS).The test results show that the light-weight lane detection algorithm with semantic segmentation can meet the real-time and accurate lane detection tasks of vehicle-mounted embedded devices.