基于高维异构信息稀疏表征与测度优化的亚米级融合定位理论方法

Sub-meter accuracy and high available indoor positioning techniques are urgently needed with the rapid development of intelligent service robots. The existing positioning solutions, like Ultra Wide-band (UWB) and Simultaneous Localization and Mapping (SLAM), which usually need the deployment of massive devices, could not meet the demands of high available positioning because of the disadvantages of high cost, high consumption of computing resource and weak environmental suitability. Therefore, multi-source fusion positioning is becoming one of the research hotspots. There are still some open challenges in multi-source fusion positioning, such as the high dimensionality and the representation disunity of heterogeneous information, and the tightly coupled fusion of heterologous information. Aiming at these problems, this project would explore the mechanism of sparse representation based on channel state information (CSI) and image scale-invariant feature transform (SIFT) keys, study the methods of bases dictionary adaptive learning and of the optimal observation matrix architectural design, and establish the dimensionality reduction computing model for CSI and image SIFT keys to realize efficient location evaluation. This project would also reveal the similarity change rules in heterogeneous parameters, propose the description methods of the optimal eigenvector to CSI and image SIFT keys fusion features, design the space measure model of heterogeneous features and the corresponding optimization strategy of model parameters, and achieve the united characterization of these two features to build a fingerprinting database with sub-meter resolution. Meanwhile, this project would study the tightly coupled fusion architecture based on the sub-meter-resolution feature, and develop high accuracy fusion positioning algorithms in order to break through the bottleneck of high available positioning and to achieve sub-meter, low-cost and high-available indoor localization, which would be validated through experiments based on the low-cost intelligent service robot platform. The research in this project would give a theoretical support to the location technologies in the applications of artificial intelligence, and provide new ideas and methods for the studies of indoor high accuracy positioning.

亚米级高可用室内定位技术是智能服务机器人等不可或缺的核心支撑。然而现有技术（如UWB、SLAM等）需布设大量设备，成本高、计算资源消耗大，无法满足高可用定位需求，发展多源融合定位理论方法成为国际学术研究的前沿和热点，但面临异构信息高维度、表征不统一与异源数据紧耦融合等科学难题。对此，本项目拟探究无线和图像高维异构信息的稀疏压缩机理，研究基字典自适应学习与观测矩阵最优结构设计方法，建立降维计算模型，大幅降低信息维度，实现高效计算；揭示异构参数相关性变化规律，提出异构信息融合的最优特征向量描述方法，设计空间测度模型与参数优化策略，实现统一的测度表征，并构建亚米级区分度特征库；研究基于亚米级区分度特征的异源数据紧耦融合高精度定位方法，突破高可用定位瓶颈，形成亚米级、低成本、高可用室内定位系统性体系。该研究不但能为面向人工智能应用的定位技术提供理论支撑，而且将为高精度室内定位技术研究提供新思路。

亚米级高可用室内定位技术是智能服务机器人等不可或缺的核心支撑，发展多源融合定位理论方法是低成本、高可用的室内定位服务的重要手段。本项目针对多源融合定位面临的异构信息高维度、表征不统一与异源数据紧耦融合等科学难题，开展了基于高维异构信息稀疏表征与测度优化的亚米级融合定位理论方法研究：在异源异构特征的稀疏表征与压缩方法方面，本项目提出了基于CSI的双层字典学习等多种算法，建立了降维计算模型，大幅降低了信息维度，实现了高效计算；在异构特征空间测度模型方面，本项目提出了LC-DNN等多种模型与算法，实现了异构信息的统一测度表征，并构建了亚米级区分度特征库，克服了在复杂环境下特征混叠冗余的问题；在异源异构数据紧耦融合稳健定位方面，本项目提出了MI-FRM等紧耦合定位方法，解决了环境复杂性带来的单一定位特征信息不完备的问题，突破了高可用定位瓶颈，形成了亚米级、低成本、高可用室内定位理论体系。最终本项目成功在端侧完成算法部署，并实现复杂室内环境下定位精度优于1米的目标。本项目的研究成果为室内定位的研究提供理论方法支撑，对促进位置服务和未来大众普及的人工智能应用的新发展具有重要意义。相关成果得到了国内外专家的高度评价，发表论文16篇（其中SCI检索论文8篇，EI检索8篇），获得发明专利6项，培养博士、硕士研究生共18名。