基于多类地标和深度强化学习轨迹融合的群智感知室内定位算法研究

Crowdsensing indoor positioning provides large-scale indoor positioning services for many emerging applications, such as smart city and intelligent transportation. which leverages crowdsourcing sensor data to avoid site surveying for indoor positioning. Because the crowdsourcing traces are normally with large quantity, short length, large heading errors, and uncertain absolute locations, the traditional GraphSLAM and unsupervised learning algorithms are difficult to achieve trace merging, and the sequential decision trace merging algorithm encounters the curse of dimensionality. Furthermore, the traditional indoor landmarks are normally with low density and low discrimination, and correspondingly difficult to indicate the relation among different traces. To handle with the aforementioned problems, first we will investigate the new method of improving the accuracy of traces. Second, we will formulate the trace merging as a high dimensional Markov decision process, and investigate how to use the deep reinforcement learning algorithm to solve it. Third, we plan to recognize the ubiquitous visible light sources and fuse with WiFi and magnetic fingerprints to investigate the new method of using multi-landmark to increase the density and discrimination of landmarks, and correspondingly improve trace merging. Finally, based on the crowdsourcing fingerprints, we will adopt particle filters to fuse fingerprinting and PDR to achieve crowdsensing positioning with high accuracy. The proposed trace merging with multi-landmark and deep reinforcement learning will improve the efficiency and accuracy of trace merging. It will be the key technology for crowdsensing positioning and promote the development of location based services for many emerging applications.

群智感知定位利用众包传感器信息，实现零人工勘测室内定位，为智慧城市和智能交通等新兴领域提供大范围室内定位。由于众包用户轨迹具有数量大、长度短、航向误差大且绝对位置未知的特点，传统GraphSLAM和无监督学习难以有效实现轨迹融合，而序列决策轨迹融合则面临维数灾难。另外，传统室内地标分布密度低且区分度低，难以准确指示轨迹关联性。针对上述问题，1. 研究提高轨迹生成精度新方法。2. 将多用户轨迹融合建模为高维马尔科夫决策过程，并探究利用深度强化学习实现轨迹融合。3. 识别室内普适性可见光光源，并融合WiFi和地磁指纹，探求如何通过多类地标提高地标密度和区分度，进而提高轨迹融合精度。4. 基于众包指纹库，通过粒子滤波融合指纹和惯性定位实现高精度群智感知定位。本项目提出的多类地标和深度强化学习轨迹融合将提高轨迹融合精度和效率，成为真实场景中群智感知定位的关键技术，推动众多新兴领域位置服务发展。

群智感知定位利用众包传感器信息，实现零人工勘测室内定位，为智慧城市、智能交通等新兴领域提供大范围室内定位。在真实众包数据采集场景下，众包用户轨迹具有数量大、长度短、航向误差大且绝对位置未知的特点，此外传统基于WiFi信号的室内定位系统易受室内多径和动态环境的影响。针对上述真实众包场景下的群智感知定位问题，项目组从轨迹获取、前端多信息融合关联识别、后端轨迹融合与地图构建三方面展开研究。在轨迹获取方面，项目组利用XGBoost和半监督学习实现了高于90%的室内运动状态识别精度，为自动化众包轨迹采集和筛选奠定了基础。在前端多信息融合关联识别方面，项目组分析了WiFi、地磁、可见光等不同信息标志，并通过多种信息融合提高室内信息标志的分布密度和区分度，准确指示轨迹内和轨迹间的关联性，从而实现了亚米级（可见光、地磁）和米级（WiFi）群智感知定位。在后端轨迹融合与地图构建方面，项目组针对WiFi、地磁和可见光等普适信号的特点，将轨迹融合问题构建为图优化问题和轨迹迭代控制问题，通过对GraphSLAM，ICP SLAM等轨迹融合算法进行优化，实现了亚米级（可见光、地磁）和米级（WiFi）的轨迹融合，从而能够基于融合的轨迹生成室内指纹地图。此外，项目组基于WiFi指纹地图特点，对迁移学习进行优化，实现了无需轨迹数据的指纹地图自动更新，解决了动态环境下地图失效问题。该项目对群智感知定位领域中上述三个关键部分，均进行了理论建模和优化，通过传感器融合技术、SLAM技术和人工智能技术，提高了群智感知定位的精度和效率，为未来高精度、高覆盖和高效率普适定位提供了解决方案。