【摘要】 为解决不同终端之间射频特性的变化问题,借助迁移学习将室内定位系统的设备异构性问题转化为领域适应性问题,提出了一种复杂度低的基于相关对齐和集成学习的室内定位算法-相关对齐定位(CALoc).该算法只需要将实时指纹与离线指纹的二阶统计特性进行对齐.这种实时定位的在线校准方法很大程度上消除了在线指纹库与离线指纹库的设备异构性.该算法无需任何耗时的深度学习重新训练过程,在线定位只需要0.11 s.实际场景的实验结果表明,CALoc与其他传统算法相比取得了显著的性能提高,定位精度平均优化提高了19%.更多还原

【Abstract】 To solve the problem of variations in radio frequency characteristics among different devices, transfer learning is applied to transform device diversity to domain adaptation in the indoor localization algorithm. A robust indoor localization algorithm based on the aligned fingerprints and ensemble learning called correlation alignment for localization(CALoc) is proposed with low computational complexity. The second-order statistical properties of fingerprints in the offline and online phase are needed to be aligned. The real-time online calibration method mitigates the impact of device heterogeneity largely. Without any time-consuming deep learning retraining process, CALoc online only needs 0.11 s. The effectiveness and efficiency of CALoc are verified by realistic experiments. The results show that compared to the traditional algorithms, a significant performance gain is achieved and that it achieves better positioning accuracy with a 19% improvement.更多还原