Unsupervised Learning Techniques for Trilateration: From Theory to Android APP Implementation

Abstract

Because the characteristics of wireless propagation channels (especially indoor channels) are too diverse and complex, the distance estimation strategy of range-based positioning techniques should adaptively change depending on the environment. In this paper, we study unsupervised learning techniques that efficiently do this without human intervention. As users simply move around an area of interest with mobile devices, the proposed method autonomously learns the characteristics of the surrounding environments and changes the ranging strategy accordingly. To this end, we use either model-based or neural network (NN)-based ranging modules for estimating the distance from neighboring anchor nodes, calculate the position of the devices using trilateration techniques, and define cost functions that indirectly evaluate the accuracy of the ranging module based on the trilateration results. Moreover, by assigning a unique trainable variable to each device, the proposed method is also able to compensate for different characteristics between devices without ground truth data. The performance of the proposed method is verified with a real-time location tracking application using received signal strength (RSS) measurements from conventional Wi-Fi access points (APs) or round trip time (RTT) measurements from APs that support the fine timing measurement (FTM) protocol. In cases where a model-based ranging module is used, the proposed method closely achieves the benchmark performance, which perfectly optimizes all the trainable variables on the test data. If NNs are adopted in the ranging module, the proposed method even outperforms the benchmark and achieves an average positioning accuracy of up to 2.397 m using RSS measurements, and up to 1.547 m using RTT measurements under the 40 MHz bandwidth configuration.