Efficient IoT Devices Localization Through Wi-Fi CSI Feature Fusion and Anomaly Detection

TL;DR

The paper tackles indoor IoT device localization by exploiting smartphone-based self-localization to enable direct IoT localization from Wi-Fi CSI. It introduces LoSEstNet to fuse CSI features across short trajectory segments for accurate LoS-AoA estimation and AnoDetNet to identify and remove anomalous CSI sequences, thereby boosting localization reliability. Localization is performed with a simple least-squares step using the refined LoS-AoA estimates and smartphone positions, achieving decimeter-level accuracy in most cases and around 1 m accuracy in 90% of trials, validated through simulations and real-world experiments with a two-element patch array. The approach also generalizes to UWB, demonstrating competitive LoS-AoA accuracy and robust anomaly filtering across different bandwidths and antenna configurations, highlighting its practicality for real-world indoor IoT deployments.

Abstract

Internet of Things (IoT) device localization is fundamental to smart home functionalities, including indoor navigation and tracking of individuals. Traditional localization relies on relative methods utilizing the positions of anchors within a home environment, yet struggles with precision due to inherent inaccuracies in these anchor positions. In response, we introduce a cutting-edge smartphone-based localization system for IoT devices, leveraging the precise positioning capabilities of smartphones equipped with motion sensors. Our system employs artificial intelligence (AI) to merge channel state information from proximal trajectory points of a single smartphone, significantly enhancing line of sight (LoS) angle of arrival (AoA) estimation accuracy, particularly under severe multipath conditions. Additionally, we have developed an AI-based anomaly detection algorithm to further increase the reliability of LoSAoA estimation. This algorithm improves measurement reliability by analyzing the correlation between the accuracy of reversed feature reconstruction and the LoS-AoA estimation. Utilizing a straightforward least squares algorithm in conjunction with accurate LoS-AoA estimation and smartphone positional data, our system efficiently identifies IoT device locations. Validated through extensive simulations and experimental tests with a receiving antenna array comprising just two patch antenna elements in the horizontal direction, our methodology has been shown to attain decimeter-level localization accuracy in nearly 90% of cases, demonstrating robust performance even in challenging real-world scenarios. Additionally, our proposed anomaly detection algorithm trained on Wi-Fi data can be directly applied to ultra-wideband, also outperforming the most advanced techniques.

Figures (13)

• Figure 1: Architecture of LoSEstNet, utilized to enhance LoS-AoA estimation accuracy by leveraging adjacent trajectory points.
• Figure 2: Architecture of AnoDetNet, designed for filtering out unreliable trajectory points in localization tasks.
• Figure 3: Illustration of the Simulation Scenario: Fixed IoT devices and smartphones on the move, demonstrating various trajectories within the annotated area (enclosed by the blue box), showcasing the dynamic interaction essential for localization accuracy testing.
• Figure 4: Comparison of Reconstruction Error ${\cal E}$ in AnoDetNet Based on LoS-AoA Estimation Errors: Error $< 15^\circ$ vs. Error $> 15^\circ$ in LoSEstNet.
• Figure 5: CDF comparison of various LoS-AoA estimation methods within the annotated area shown in Fig. \ref{['fig:floorplan']}.