基于多层次特征描述的高分辨率遥感影像交通目标检测技术

项目名称： 基于多层次特征描述的高分辨率遥感影像交通目标检测技术

项目编号： No.41301453

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 天文学、地球科学

项目作者： 陶超

作者单位： 中南大学

项目金额： 25万元

中文摘要： 从高分辨率遥感影像中自动检测常见的交通类目标(车辆和船只），无论在民用还是军事领域都有着广阔的应用前景。然而在交通目标描述过程中，现有方法多使用人工挑选的低层次视觉特征，而较少关注高层次的目标语义特征，导致目标解译过程中所使用的特征区分性不强，所需要的特征不完备，进而影响后续的目标检测与识别。鉴于此，本课题拟在充分考虑高分辨率遥感影像交通目标特点的基础上，研究多层次的交通目标特征建模方法：(1)基于多核学习的目标低层次视觉特征建模方法；(2)基于概率图模型的目标高层次语义特征建模方法。然后根据特征描述的层次性，建立"低层次视觉特征-候选目标初提取-高层次语义特征-目标提取结果精炼"逐层关联的目标检测机制，为构建实用、有效的高分辨率遥感影像交通目标检测系统提供科学依据。研究成果将能为高分辨率遥感影像智能解译提供理论与关键技术支持，为我国智能交通监控系统的研制提供技术支撑。

中文关键词： 交通目标识别；低层次视觉特征；高层次语义特征；高分辨率遥感影像；智能解译

英文摘要： With the development of satellite and remote sensing technologies, the effective extraction of traffic objects (e.g. car and ship) from high-resolution remote sensing image has become imperative as applied in both civilian and military field. However, as current research on feature modelling of traffic object is not deep enough, it is challenging to define highly discriminative and robust features to describe them, and further separate them from other objects. Consequently, the main goal of this project is to model multilevel feature representation (e.g. low-level visual feature and a high-level semantic feature) for traffic objects, by using multiple kernel learning and probabilistic graphical model. With the hierarchical description of object features, it is devoted to build a hierarchical traffic object detection framework as "low-level visual feature - candidate objects extraction - high level semantic feature - object detection refinement". The achievement of this project can be used as theoretical basis and technical support for high-resolution remote sensing image understanding, and applied in fields like intelligent traffic supervision.

英文关键词： Traffic object detection；Low-level visual feature；High-level semantic feature；High-resolution Remote Sensing Image；Intelligent interpretation