TL;DR
This paper presents a novel indoor location-based service (LBS) system for smartphones that fuses Inertial Navigation System (INS) data, Wi-Fi fingerprinting, and indoor map information. By employing a two-layer Kalman/Particle filter architecture, the system achieves accurate, continuous, and low-cost indoor positioning, significantly reducing the drift inherent in standalone INS.
What Problem Does This Paper Address?
Accurate and reliable indoor localization remains a significant challenge, particularly for mobile devices. While GPS excels outdoors, it's ineffective indoors. Existing indoor LBS solutions often rely on infrastructure like Wi-Fi access points or Bluetooth beacons, which can be costly to deploy and maintain, or are sensitive to environmental changes. Inertial Navigation Systems (INS) using low-cost MEMS sensors offer a self-contained solution but suffer from rapid error accumulation due to sensor integration. This paper tackles the problem of creating a robust, accurate, and cost-effective indoor positioning system for smartphones by intelligently combining these disparate data sources.
Methods
The proposed system integrates three core modules:
- INS Module: Utilizes smartphone's accelerometers and gyroscopes to provide a primary, though error-prone, position estimate through Strapdown Inertial Navigation System (SINS). Non-holonomic constraints (NHC) are applied to limit velocity errors.
- Wi-Fi Module: Employs Wi-Fi fingerprinting, a common technique that matches received signal strengths (RSSI) from access points against a pre-built radio map. The Weighted K-Nearest Neighbor (WKNN) algorithm is used for matching. This module is only active when Wi-Fi signals are available.
- Map Module: Leverages indoor map information to constrain the estimated position.
The key innovation lies in the two-layer Kalman/Particle filter cascade structure:
- Lower Layer (Kalman Filter): This layer fuses the INS position with Wi-Fi derived positions (if available) and NHC. The Kalman filter is chosen for its efficiency and ability to handle high-rate sensor measurements, correcting the preliminary INS solution.
- Upper Layer (Auxiliary Particle Filter - APF): This layer incorporates indoor map information. The APF is used because it can more effectively handle the non-linearities and uncertainties associated with map-matching and is less prone to particle impoverishment than traditional particle filters. Map information is used in the measurement update phase, specifically through a "wall-cross" method, where particles intersecting with walls are penalized.
This cascaded approach aims to leverage the strengths of each filter type and data source while mitigating their weaknesses, all without significantly increasing computational burden.
Datasets
The paper describes experiments conducted on a Samsung Galaxy Note 4 smartphone. The specific datasets used for evaluation are not explicitly detailed as pre-existing benchmarks. Instead, the authors generated their own data by recording sensor measurements and Wi-Fi RSSI values during a 25-minute walk along a predefined trajectory within the Energy Environment Experiential Learning (EEEL) building at the University of Calgary. The reference trajectory was obtained separately.
Key Insights and Ideas
- Synergistic Fusion is Crucial: Standalone INS is insufficient due to rapid error accumulation. Wi-Fi fingerprinting can be noisy and exhibit "jumping points." Combining them, and then further constraining with map data, is essential for robust performance.
- Two-Layer Filtering for Efficiency and Accuracy: The Kalman filter efficiently handles high-frequency INS data and Wi-Fi fusion, while the Particle Filter (specifically APF) excels at incorporating the more complex, non-linear map constraints without excessive computational cost.
- Map-Aiding as a Powerful Constraint: Indoor maps, even with limited architectural detail (as noted in the experiment environment), can significantly improve the accuracy and stability of the estimated trajectory by preventing the system from drifting into invalid areas (e.g., through walls).
- Cost-Effectiveness: By utilizing ubiquitous Wi-Fi signals and readily available smartphone sensors and maps, the proposed system offers a low-cost solution for indoor LBS.
Key Results
The experimental results demonstrate the effectiveness of the proposed method:
- INS alone: Produces a highly inaccurate and drifting trajectory, even with NHC constraints.
- Wi-Fi fingerprinting: Shows improvement over INS alone but suffers from noticeable "jumping points" and instability.
- INS/Wi-Fi integrated (Kalman Filter): Provides a better lower-filter solution than either INS or Wi-Fi alone, but still exhibits significant drift.
- Map/INS/Wi-Fi integrated (APF): The final integrated system, using map aiding, shows a trajectory that "almost matches the reference trajectory," indicating a substantial improvement in accuracy and stability. The paper mentions that combining with a "Map matching method" further improves the RMS error by about 1 meter, though this specific method isn't detailed in the core algorithm description.
Limitations and Future Work
- Limited Map Detail: The experimental environment (a hall area) had limited architectural map information, which the authors note could have provided stronger constraints. Future work could explore more detailed map data.
- Wi-Fi Availability: The system's performance is dependent on the availability of Wi-Fi signals. When Wi-Fi is unavailable, the system relies solely on INS and map information.
- Computational Burden: While the system is designed to be computationally efficient, the paper doesn't provide specific metrics on processing time or resource usage on the smartphone.
- Map Matching Method: The mention of an additional "Map matching method" improving results suggests potential for further optimization beyond the described map-aiding approach. Future work could delve deeper into advanced map-matching techniques.
Summary updated