TL;DR
This paper presents a novel two-layer Extended Kalman Filter (EKF) and Auxiliary Value Particle Filter (AVPF) system that integrates Inertial Navigation System (INS) data with Wi-Fi fingerprinting and indoor map information. The proposed method significantly improves indoor localization accuracy and continuity, offering a low-cost and self-contained solution for location-based services.
What Problem Does This Paper Address?
Indoor localization remains a significant challenge, particularly for commercial applications where cost and reliability are paramount. While Inertial Navigation Systems (INS) offer continuous positioning, they suffer from rapid error accumulation due to sensor drift. Existing solutions often rely on external infrastructure (like GPS, which is unavailable indoors) or are computationally expensive. This paper tackles the problem of achieving accurate, continuous, and low-cost indoor localization by effectively fusing data from readily available sources: smartphone inertial sensors (IMU), ubiquitous Wi-Fi signals, and existing indoor maps. The core challenge is to intelligently combine these complementary data sources to overcome the limitations of each individual technology.
Methods
The paper proposes a two-layer EKF/AVPF structure for integrating various data sources.
- Lower Layer (EKF): This layer focuses on fusing IMU data with Wi-Fi positioning and non-holonomic constraints (NHC). NHC helps correct for the unobservable heading error in INS. When Wi-Fi signals are available, they are used to further refine the INS solution, improving accuracy and reducing the computational load on the upper layer. If Wi-Fi is unavailable, the system falls back to using indoor map information.
- Upper Layer (AVPF): This layer primarily utilizes indoor map information to constrain the primary Wi-Fi/INS solution. An Auxiliary Value Particle Filter (AVPF) is employed, which is an enhancement over traditional Particle Filters. AVPF addresses particle impoverishment by incorporating measurements into the importance density function and uses a Pedestrian Dead Reckoning (PDR) model for step and heading updates.
- Map Aiding (MA) and Map Matching (MM): These methods are integrated into the AVPF layer. MA uses the indoor map to constrain the estimated solution, while MM projects the estimated position onto the closest link in the digital map network to further refine accuracy. The paper specifically uses a point-to-point map-matching method.
Datasets
The paper does not explicitly mention using pre-existing datasets for training. Instead, it describes experiments conducted in two real-world indoor environments at the University of Calgary: the basement floor of the Engineering (ENG) building and the first floor of the Energy Environment Experiential Learning (EEEL) building. The "reference trajectories" for evaluation were generated using a "heading lock" PDR method, which assumes a locked walking heading according to the indoor map and uses corners/intersections as landmarks for correction.
Key Insights and Ideas
- Complementary Fusion: The core insight is that by intelligently fusing IMU, Wi-Fi, and map data, the weaknesses of each individual sensor can be mitigated, leading to a robust and accurate system.
- Two-Layered Approach: The EKF/AVPF structure effectively balances computational efficiency with accuracy. The EKF handles the high-frequency sensor data, while the AVPF leverages map information at a lower update rate, reducing overall computational burden.
- AVPF for Map Integration: The use of AVPF is crucial for effectively incorporating map information, overcoming limitations of traditional particle filters in handling diverse sensor characteristics and motion dynamics.
- Wi-Fi as a Cost-Effective Aiding Source: Leveraging ubiquitous Wi-Fi signals, which require no additional infrastructure, is a key factor in achieving a low-cost solution.
- Map Matching Refinement: The addition of map matching as a final step significantly refines the trajectory, bringing it closer to the true path.
Key Results
The paper demonstrates significant improvements in localization accuracy through experimental results.
- RMS Error Reduction:
- INS/NHC alone: RMS error of 23.1 m (ENG) and 16.8 m (EEEL).
- INS/Wi-Fi integrated: RMS error reduced to 7.2 m (ENG) and 7.3 m (EEEL).
- Map Aiding (MA)/INS/Wi-Fi: Further improved accuracy, with RMS errors of 4.7 m (ENG) and 5.7 m (EEEL).
- Map Matching + Map Aiding (MM+MA)/INS/Wi-Fi: The proposed final method achieved RMS errors of 3.0 m (ENG) and 4.8 m (EEEL).
- Low Computational Burden: The proposed method requires fewer particles compared to traditional PF, indicating a lower computational burden. For example, 1000 particles were used for the ENG test, and 2000 for the EEEL test.
- Qualitative Improvement: Visualizations of the trajectories show that the proposed method's estimated paths closely match the reference trajectories, especially after incorporating map matching.
Limitations and Future Work
- Limitations:
- The paper acknowledges that in areas with sparse Wi-Fi signals or limited architectural map information (e.g., large halls), the system's accuracy might be affected.
- The "heading lock" PDR method used for generating reference trajectories might introduce some bias in the ground truth.
- Future Work:
- The authors plan to test the algorithm in multi-floor environments to provide continuous navigation across different levels.
- They aim to program the algorithm onto a smartphone platform for real-time, on-device implementation and conduct real-time experiments to further demonstrate its low computational burden.
Summary updated