复杂交通场景下基于深度迁移学习的车辆识别方法研究

Vehicles in complex traffic scene appear large interclass difference and contaminated property. The classifier trained with limited sample is not able to satisfy the time-vary feature of on-board platform and diversity of complex scene. Thus the accuracy of vision based vehicle recognition is still low which constrain the development of intelligent vehicles. Based on deep learning theory, this project tends to research on new hierarchical visual feature learning and classifier building model focus on complex traffic images. The low-level feature extraction method based on image feature multiple nonlinear isolated subspace distribution assumption is explored. The simplified network structure for resource constrained on-board platform is established. After that hierarchical feature learning model with sparse constraint is proposed to construct high discriminating visual feature representation for vehicle target. Based on this, this project further introduces transfer learning theory to deep network model. Under the framework, proposing a new sample selection and labeling methods with label confidence, exploring bottom-up low level features unsupervised transfer model, searching top-down label confidence based high classifier surface transfer method, so that to archive the transfer learning for deep network classifier and improve the generalization ability of vehicle classification under complex traffic environment. At last, the whole proposed algorithm will be verified in on-board platform. Research of this project is able to satisfy high accuracy requirement of vehicle recognition for intelligent vehicles in complex traffic scene and is with high theoretical and practical value.

复杂交通场景下的车辆呈现较大类内差异性和受污染性，有限样本集训练所得分类器难以满足车载平台的动态特性和交通场景的多样性，视觉车辆识别准确性低，制约了智能车辆的发展。本项目基于深度学习理论，研究面向复杂交通图像的新型分层视觉特征学习方法与分类器构建模型，探索基于图像特征多非线性独立子空间分布假设的底层特征抽取方法，建立面向车载资源受限平台的精简化深度网络结构，提出基于稀疏约束的逐层特征计算模型，揭示车辆目标的高判别性视觉特征表达；进一步将迁移学习理论引入深度网络模型，提出赋予标签置信度的新场景样本标记方法，探索自底向上的底层特征无监督迁移机制，寻求自顶向下引入标签置信度的高层分类面迁移方法，实现深度网络分类器在动态场景下的迁移，提升复杂交通场景下车辆分类器的泛化能力；最后基于车载平台进行算法验证。研究成果可满足智能车辆复杂交通场景下高准确性车辆识别的需求，具有重要的理论意义与实用价值。

复杂交通场景下的车辆呈现较大类内差异性和受污染性，有限样本集训练所得分类器难以满足车载平台的动态特性和交通场景的时变性、多样性，视觉车辆识别准确性低，制约了智能车辆的发展。本项目引入视觉显著性机制，提出基于多尺度深度网络的显著性区域提取方法，并将该方法应用到道路图像的前景感兴趣区域提取中，效果优异。构建了面向复杂交通图像的新型分层视觉特征学习方法与分类器构建模型，研究了基于图像特征多非线性独立子空间分布假设的底层特征抽取方法，建立面向车载资源受限平台的精简化深度网络结构，设计了基于稀疏约束的逐层特征计算模型，在KITTI等国际主流数据库的测试结果标明了所提算法在复杂交通场景下具有优秀的车辆检测性能。将迁移学习理论引入深度网络模型，提出赋予标签置信度的新场景样本标记方法，构建自底向上的底层特征无监督迁移机制和自顶向下引入标签置信度的高层分类面迁移方法，实现深度网络分类器在动态场景下的迁移，大幅提升车辆分类器在不同场景下的泛化能力。基于权威公开测试集和车载平台进行算法验证，结果表明所提算法可满足智能车辆复杂交通场景下高准确性车辆识别需求，具有重要理论意义与实用价值。