多媒体传感器网络中目标多分类问题研究

With rapid improvements and miniaturization in hardware, sensor nodes equipped with acoustic and visual information collection modules promise an unprecedented opportunity for target surveillance applications. Automatic target classification is an important fundamental function of surveillance applications. However, existed works about target classification in multimedia sensor networks mainly focus on the bianry classification. There is few research on the multi-classification which is more applicable than the bianry classification. Towards this end, this project plans to investigate the multi-class classification problem for multimedia sensor networks.Our research will focus on the following aspects: sensing model and coverage analysis, lightweight target detection and feature extraction, multi-classifier generation and division, and optimal deployment of sub-classifier set. Our goal is to buld a series of methods for high-efficient ultilizing multiple feaures and accuract classifying the target in multimedia sensor networks. Furthermore, our methods also need to take the balance between the resource limitation of multimedia sensor nodes and the high-complexity of multi-classification task into account. Finally, based on a multimedia sensor network, this project will construct a prototype of multi-classification system to validate and evaluate our research results.

随着硬件能力的提升和小型化技术的发展，装配有音视频信息采集模块的多媒体传感器节点将被广泛应用于目标监控应用中，并展现出前所未有的优势。对于目标监控应用而言，对目标进行自动识别分类是监控应用的重要基础，也是监控智能化的体现。目前，对多媒体传感器网络中的目标分类研究仅仅停留在简单的二分类上面，鲜有针对实用性更高的多分类的研究成果。因此，本课题将研究多媒体传感器网络中的目标多分类问题。本课题从多媒体传感器节点感知模型与面向分类的覆盖分析、轻量级目标检测与特征提取方法、基于多种特征的多分类器构造与划分方法、子分类器集优化部署四个方面展开研究，通过一整套方法实现高效利用多媒体传感器节点提取的多种特征，准确地对目标进行多分类，并解决多媒体传感器节点的有限能力与多分类的高计算复杂度之间的矛盾。最后通过构建基于多媒体传感器网络的目标多分类原型系统，对课题取得的理论成果与关键技术方法进行验证。

课题组在本基金项目支持下对多媒体传感网络感知覆盖分析、目标分类与定位、协同计算与参与激励、视频机会传输、以信息为中心互联五个方面进行了系统研究，取得了系列成果，在相关领域国际期刊和会议上发表论文20篇（SCI论文10篇），其中发表在IEEE TC/TMC、ACM MM共计3篇CCF推荐A类期刊或会议论文。获得授权发明专利2项，申请发明专利10项，包括2项美国专利。研制了多媒体传感器节点与无线多模网关节点平台，实现了多媒体传感器网络协同识别系统，以及信息为中心多媒体传感网系统，并在长城保护、城市交通等方面进行了应用。成果有效解决了多媒体传感器节点的有限能力与多分类的高计算复杂度之间的矛盾，为多媒体传感器网络中的目标监控应用提供基础理论与关键技术支持，并为物联网的进一步研究打下了基础。