利用機器人採樣實現類神經網路之地磁場室內定位

由於在室內環境收不到GPS訊號，因此有許多室內定位的技術被發展出來。這些定位技術如基於慣性感測元件、聲音訊號、可見光、無線訊號等等。在我們的實作中，我們使用地磁場搭配深度學習之類神經網路來學習與位置相關的特徵。現有的文獻大多還是需要花費人力來收集這些地磁場的資料並利用線性內插來增加資料集。然而，我們的實驗顯示出用內插產生的資料與真實量測之資料量存在著差異性，因為地球磁場值並不總是線性的。因此，我們使用機器人搭載著手機在室內空間收集較精細的地磁場資料。我們使用這些資料集來訓練我們系統的兩個定位模型:深度類神經網路以及循環類神經網路。除此之外，我們也利用單點磁力資料來合成數條的磁力軌跡，藉此增加循環類神經網路模型的資料量。從實驗結果顯示，我們提出之方法的定位誤差比先前的研究來得更加提升。

關鍵字

室內定位；地球磁場；物體追蹤；類神經網路

並列摘要

Since GPS signal is not available in indoor environments, a lot of indoor localization technologies have been proposed based on inertial sensors, audio signals, visible light, wireless signals, etc. In this work, we consider using magnetic fields integrated with deep learning neural networks to learn location-related features. Existing works try to collect magnetic field data by human and use interpolation to increase dataset. However, our experiments show that the data generated by interpolation is usually different from the ground truth because magnetic fields are not always linear. Therefore, we dispatch a robot carrying a smartphone to collect dataset at a fine resolution. We use these collected data to train our two localization models: deep neural network (DNN) and recurrent neural network (RNN). Besides, we augment our RNN training dataset by synthesizing single point magnetic data to form magnetic trajectories. Field trials are conducted, which validate that our approach outperforms previous work.