车载全景序列图像的特征匹配方法研究

项目名称： 车载全景序列图像的特征匹配方法研究

项目编号： No.61472267

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 胡伏原

作者单位： 苏州科技大学

项目金额： 78万元

中文摘要： 车载全景序列图像在构建智慧城市基础地理信息数据库中发挥着重要作用，目标匹配尤其是点匹配是其深度应用的核心内容之一。车载相机的快速运动及其复杂场景易导致点特征难以准确表示、点匹配精度低以及匹配速度慢等问题。因此,项目研究内容包括：①考虑视角变化和场景深度不一致对点准确表示的影响，结合场景结构研究自适应分数阶微分，探索目标区域准确表示机理。②在深度学习框架下，结合流形学习设计网络拓扑结构和学习算法，提取具有语义特点的结构化特征，提高目标表示准确性。③考虑相似纹理结构、运动目标和相机运动等情况对匹配精度影响，将相机运动和场景结构特性引入能量优化函数，自主构建未知结构概率图模型，降低匹配错误率；设计层次化低维子空间结构图并优化模型，提高匹配效率。④设计适合全景图像应用的点表示和匹配方法,优化应用系统。研究结果可望提高任意点匹配精度和相机参数估计的鲁棒性，为智慧城市建设中实景化深度应用奠定基础。

中文关键词： 深度学习；概率图模型；分数阶微分；特征提取；流形学习

英文摘要： Panoramic image sequences based on the mobile vehicles play a critical role in geographic information database construction for smart cities. Object matching, especially point matching, is one of the key contents for its wide applications. However, the performance in terms of feature representation, matching accuracy and efficiency deteriorates due to complex scenes and fast moving cameras. The research will focus on four parts: (1) Considering scene depth inconsistency and view change, we study fractional differential with scene graphs, and find mechanism of accurate expression. (2) Under deep learning framework,we extract structural features with semantic characteristic via designed network topology and algorithms based on manifold learning to improve the performance of object representation.(3) With consideration of matching stability affected by similar texture structure, moving object and camera movement, we introduce scene graphs and camera movement to energy optimization function and construct unknown structure probabilistic graphical models, leading to error reduction for matching. Through optimizing parameter and designing hierarchy subspace structural graph, matching efficiency is improved. (4) In order to optimize the application system，the proper method for point representation and matching in panoramic image sequences is designed.The reseach findings are expected to improve matching accuracy for any points and robustness for vehicle camera parameter estimation. Meanwhile, it lays the solid foundation for wide applications of panoramic image sequences in the smart cities construction.

英文关键词： Deep Learning;Probabilistic Graph Models;Fractional differential;Feature Extraction;Manifold Learning