摘要
无人驾驶车辆在结构化道路中需要通过车道线判断自身位置,为提高其检测的实时性与准确性,提出一种利用改进分割网络(segmentation network,SegNet)算法与连通域约束相结合的方法实现车道线检测识别。将对称的SegNet算法改为非对称结构对车道线作逐像素提取:利用卷积与池化提取车道线特征,摒弃传统的车道线聚类过程,利用二值化图像结合连通域约束与关联对车道特征点进行分类,最后对相同类别的车道特征点进行车道线拟合。将该改进的SegNet算法在CULane数据集和TuSimple数据集上进行了训练与测试,结果表明:对车道分割准确、实时处理能力优秀,检测识别效果优于传统SegNet网络算法,其平均检测精度为94.60%,每帧检测耗时提升53 ms。
The position of the unmanned vehicle needs to be identified through the lane line on the structured road.In order to improve the real-time and accuracy of the detection,a method combining the improved SegNet algorithm and connected domain constraints was proposed to realize the detection and recognition of lane lines.The symmetric SegNet algorithm was changed to an asymmetric structure to extract the lane lines pixel by pixel.Firstly,both the convolution and pooling were used to extract lane line features.The traditional lane line clustering process was discarded,while the binarized image was used to combine connected domain constraints and associations to classify lane feature points.Finally,the lane feature points of the same category were fitted by lane lines.The algorithm was trained and tested on the CULane dataset and the TuSimple dataset.The results are with accurate lane segmentation and excellent real-time processing capabilities.Its effect of detection and recognition is better than the traditional SegNet algorithm with its average detection accuracy of 94.60%,and the detection time per frame is improved by 53 ms.
作者
邓天民
王琳
杨其芝
周臻浩
DENG Tian-min;WANG Lin;YANG Qi-zhi;ZHOU Zhen-hao(College of Traffic&Transportation,Chongqing Jiaotong University,Chongqing 400074,China)
出处
《科学技术与工程》
北大核心
2020年第36期14988-14993,共6页
Science Technology and Engineering
基金
国家自然科学基金(51678099)
重庆市科学技术委员会科技人才培养计划(CSTC2013KJRC-QNRC0148)。
关键词
无人驾驶
辅助驾驶
深度学习
卷积神经网络
语义分割
SegNet网络
车道线检测
driverless
assisted driving
deep learning
convolutional neural network
semantic segmentation
SegNet algorithm
lane line detection
