Abstract
We propose a novel method for improving Wi-Fi positioning accuracy while reducing the energy consumption of mobile devices. Our method presents three contributions. First, we jointly and intelligently select the optimal subset of access points for positioning via maximum mutual information criterion. Second, we further propose local discriminant embedding algorithm for nonlinear discriminative feature extraction, a process that cannot be effectively handled by existing linear techniques. Third, to reduce complexity and make input signal space more compact, we incorporate clustering analysis to localize the positioning model. Experiments in realistic environments demonstrate that the proposed method can lower energy consumption while achieving higher accuracy compared with previous methods. The improvement can be attributed to the capability of our method to extract the most discriminative features for positioning as well as require smaller computation cost and shorter sensing time.
Introduction
Indoor positioning has become a critical component of Location-Based Services (LBS), yet it remains a persistent challenge because GPS signals struggle to penetrate buildings. The industry standard is "Wi-Fi fingerprinting," where a mobile device measures Received Signal Strength (RSS) from various Access Points (APs) to estimate its location. However, two major hurdles remain: high energy consumption from constant scanning and the inherent complexity of noisy, nonlinear Wi-Fi signals caused by multipath effects and environmental dynamics. Existing solutions often rely on simple linear models or basic AP selection strategies that fail to capture the complex correlations between signals.
The LLDE-APS Framework
To address these challenges, we propose LLDE-APS (Localized Local Discriminant Embedding and AP Selection), a three-stage framework designed to optimize both accuracy and efficiency. The system architecture is divided into an offline training phase and an online positioning phase.
1. Fuzzy C-Means (FCM) Clustering
Instead of treating an entire building as one giant map, the system uses FCM to divide the area into smaller "sub-regions." This localization reduces the search space for the positioning algorithm and makes the input signal data more compact, significantly lowering the computational burden on the mobile device.
2. Joint Mutual Information (JointMMI)
Rather than selecting APs based solely on individual signal strength, LLDE-APS employs a "joint" approach. By using a maximum mutual information criterion, the system identifies a subset of APs that provides the most informative and least redundant data. This ensures that the most discriminative signals are used for positioning.
3. Local Discriminant Embedding (LDE)
LDE is the core innovation for handling signal complexity. It assumes that high-dimensional Wi-Fi data resides on a simpler, curved manifold. LDE extracts features that keep similar locations close together in the feature space while pushing different locations apart. While the training process is sophisticated, the resulting online transformation is a simple linear projection, ensuring rapid execution on mobile hardware.
Experimental Evaluation
The researchers evaluated LLDE-APS in two real-world settings: a complex indoor office environment (25.3m x 24.7m) with significant obstructions and an open hall (20m x 16m) with line-of-sight conditions. The method was compared against industry standards such as Weighted K-Nearest Neighbor (WKNN), Maximum Likelihood (ML), and traditional feature extraction methods like PCA and LDA.
Key Results and Insights
- Significant Energy Savings: The biggest energy saving comes from reducing "sensing time"—the duration the Wi-Fi chip must stay active. LLDE-APS achieved high accuracy with just 1 second of sensing, representing an 80% reduction compared to traditional methods requiring 5 seconds.
- Computational Efficiency: The framework reduced the expected computation cost by 90.2% compared to standard WKNN, primarily due to the localized model and reduced feature dimensionality.
- Improved Accuracy: In the complex office environment, LLDE-APS achieved 73.7% accuracy within 2 meters, significantly outperforming WKNN (57.5%) and ML (59.8%).
- Feature Compactness: The system successfully reduced location features to just 3 dimensions in most clusters, compared to the 13+ dimensions typically required by other selection methods.
Conclusion
LLDE-APS demonstrates that more data is not always better; by intelligently selecting a subset of APs and applying manifold-based feature extraction, we can achieve superior positioning accuracy with a fraction of the energy and computational cost. While the offline training phase is computationally intensive, the online performance gains make it a highly viable solution for modern mobile devices. Future work may explore automating the re-mapping process to account for dynamic environmental changes.
Summary updated