基于深度卷积神经网络的场景自适应道路分割算法

摘要

The existed machine learning based road segmentation algorithms maintain obvious shortage that the detection effect decreases dramatically when the distribution of training samples and the scene target samples does not match. Focusing on this issue, a scene adaptive road segmentation algorithm based on Deep Convolutional Neural Network (DCNN) and auto encoder is proposed. Firstly, classic Slow Feature Analysis (SFA) and Gentle Boost based method is used to generate online samples whose label contain confidence value. After that, using the automatic feature extraction ability of DCNN and performing source-target scene feature similarity calculation with deep auto-encoder, a composite deep structure based scene adaptive classifier and its training method are designed. The experiment on KITTI dataset demonstrates that the proposed method outperforms the existed machine learning based road segmentation algorithms which upgrades the detection rate on average of around 4.5%.%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GentleBoost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%。