基于深度卷积神经网络的场景自适应道路分割算法

发布时间：2018-03-03 13:56

本文选题：道路分割　切入点：场景自适应　出处：《电子与信息学报》2017年02期 　论文类型：期刊论文

【摘要】：现有基于机器学习的道路分割方法存在当训练样本和目标场景样本分布不匹配时检测效果下降显著的缺陷。针对该问题,该文提出一种基于深度卷积网络和自编码器的场景自适应道路分割算法。首先,采用较为经典的基于慢特征分析(SFA)和Gentle Boost的方法,实现了带标签置信度样本的在线选取;其次,利用深度卷积神经网络(DCNN)深度结构的特征自动抽取能力,辅以特征自编码器对源-目标场景下特征相似度度量,提出了一种采用复合深度结构的场景自适应分类器模型并设计了训练方法。在KITTI测试库的测试结果表明,所提算法较现有非场景自适应道路分割算法具有较大的优越性,在检测率上平均提升约4.5%。
[Abstract]:The existing road segmentation methods based on machine learning have the defect that the detection effect drops significantly when the distribution of training samples and target scene samples mismatch. In this paper, a scene adaptive road segmentation algorithm based on deep convolution network and self-encoder is proposed. Firstly, the classical methods based on slow feature analysis (SFAs) and Gentle Boost are used to realize the on-line selection of labeled confidence samples. Based on the feature extraction ability of DCNN (depth convolution neural network), the feature similarity measurement in source-target scene is presented with feature self-encoder. A scene adaptive classifier model with compound depth structure is proposed and a training method is designed. The test results in KITTI test library show that the proposed algorithm is superior to the existing non-scene adaptive road segmentation algorithm. The average increase in the detection rate was about 4.5%.
【作者单位】：江苏大学汽车与交通工程学院;江苏大学汽车工程研究院;克莱姆森大学汽车工程系;
【基金】：国家自然科学基金(U1564201,61601203,61573171,61403172) 中国博士后基金(2014M561592,2015T80511) 江苏省重点研发计划(BE2016149) 江苏省自然科学基金(BK20140555) 江苏省六大人才高峰项目(2015-JXQC-012,2014-DZXX-040)~~
【分类号】：TP391.41;TP18

【参考文献】