面向未知环境探索的类别独立对象检测

To execute tasks in unexplored environments is the requirement of high-level intelligent systems. However, these systems often do not have all category priors of all objects in such environment. On the other hand, traditional object detection methods cannot efficiently detect all interested object in unexplored environment. This project focuses on the method of efficiently detecting objects in unexplored environment, by which the self-adaption of intelligent systems can be improved greatly. For precisely detecting all interested objects in unexplored environment, this project proposes a theoretical framework, which integrates multimodal data, computes visual saliency that is based on region consistency, multimodal oversegments and multimodal boundary statistics, as well as the cognition theory of human. In this project, we plan to develop a novel high-order cross modal co-segmentation random field model to detect category-independent objects in unexplored environments. By comparison with other state-of-the-art category-dependent object detection methods and category-independent objectness ranking methods, this framework has the following advantages: 1) it is unnecessary to know the category priors when detecting objects, 2) it accurately localizes objects' position without sampling any extra regions as candidates, 3) it can distinguish overlapping objects and precisely output objects' boundaries. Therefore, the research work in this project has strong theory and practicalness, the feasibility of related methods has been proved. This project is expected to produce some ideal results.

在未知环境中执行任务是高级智能系统的需求。但是这种系统经常不存在其中全部对象的类别先验信息，而传统对象检测方法并不能有效地检测出未知环境下所有感兴趣对象。本项目研究在未知环境中有效检测对象的方法，以提高系统的自适应能力。为了精确检测未知环境中感兴趣对象，项目提出一种理论框架，通过融合多个模态的数据，计算基于区域一致性的多模态显著性，过分割和多模态边缘信息，并结合人类认知理论。项目设计一种高阶交叉模态共分割模型进行类别独立的对象检测，有别于传统的对象检测和类别独立对象性排序算法相比，该理论框架：（1）检测对象时无需知道所检测的对象的类别；（2）无需采样大量候选区域，直接精确地定位对象位置；（3）能够精确区分重叠对象，并精确给出对象轮廓。因此，本项目的研究内容具有较强的理论性和实用性，并且充分前期研究的基础上，论证了相关方法的可行性，预期可产生较理想的成果。

本项目主要研究一种可以在未知环境中能够有效检测对象的方法，可以提高系统的自适应能力。主要研究内容包括通过融合多个模态的数据，计算多模态显著性，利用过分割和多模态边缘信息作为特征，构建了一种高阶交叉模态共分割模型进行类别独立的对象检测。该方法具有如下优势：1）检测对象时无需知道所检测的对象的类别；2）无需采样大量候选区域，直接精确地定位对象位置；3）能够精确区分重叠对象，并精确给出对象轮廓。该方法可应用于：1）通过引入多模态数据提取更为精确的显著性；2）能够较为准确检测类别独立对象，为后续对象识别工作提供便利；3）为系统进行未知环境识别提供了一种新的物体检测方法。