TL;DR
This paper introduces MapSentinel, a non-intrusive indoor tracking system that leverages WiFi and ultrasonic sensors, crucially incorporating "space use knowledge" from building floorplans. By understanding how people move differently in corridors versus open areas, MapSentinel significantly improves tracking accuracy over traditional methods.
What Problem Does This Paper Address?
Accurate real-time indoor tracking is essential for smart buildings, enabling applications like optimized lighting and ventilation. However, existing systems often require occupants to carry specialized devices or install software on their phones, leading to issues with user adoption and battery drain. While some systems use floorplans to prevent unrealistic trajectories (e.g., walking through walls), they haven't effectively utilized the functional information within those floorplans – how people actually move in different types of spaces (like corridors vs. open areas). This paper aims to bridge this gap by integrating this "space use knowledge" into a non-intrusive tracking system.
Methods
MapSentinel employs a two-pronged sensing approach:
- Non-Intrusive Sensing:
- WiFi Access Points (APs): Modified to detect Received Signal Strength (RSS) from occupants' devices without requiring special apps. This provides broad coverage.
- Ultrasonic Sensors: Deployed in specific zones to provide precise, short-range distance measurements, acting as a calibration mechanism for WiFi.
- Information Fusion Framework:
- Context-Dependent Kinematic Models (CDKM): This is the core innovation. Instead of a single motion model, MapSentinel uses different models for different "contexts" derived from the floorplan (e.g., "Free Space," "Constrained Space" like corridors, "Static Space" like cubicles). This allows the system to predict movement more realistically based on the environment.
- Probabilistic Sensor Measurement Models (PSMM): Standard models are used to represent the uncertainty in WiFi and ultrasonic sensor readings.
- Factor Graph: A graphical model is used to represent the dependencies between states (position, velocity, context) and measurements.
- Context-Augmented Particle Filter (CAPF): An extension of the standard particle filter designed to handle the cyclic dependencies introduced by the context variable. It recursively predicts and updates particle states, incorporating the context information. The "oracle" for context prediction uses a K-Nearest Neighbors approach on WiFi fingerprints.
Datasets
The paper describes an experimental setup in a 1000 m² office environment at the Singapore-Berkeley Building Efficiency and Sustainability in the Tropics (SinBerBEST) facility. This environment included WiFi routers and ultrasonic stations. The paper does not mention using pre-existing public datasets for training or evaluation; instead, it relies on data collected from this specific testbed.
Key Insights and Ideas
- "Space Use Knowledge" is Powerful: The most significant insight is that understanding how people move in different types of spaces (corridors, open areas, cubicles) is a crucial, underutilized source of information for indoor tracking. This goes beyond simply knowing where walls are.
- Non-Intrusiveness is Key for Adoption: By avoiding the need for occupants to carry special devices or install apps, MapSentinel significantly increases its potential for real-world deployment.
- Contextual Motion Models Improve Predictions: The context-dependent kinematic models allow for more realistic predictions of occupant movement, reducing the likelihood of impossible trajectories and improving overall accuracy.
- Fusion is More Than Just Combining Sensors: The paper demonstrates that effectively fusing sensor data with contextual map information, rather than just combining sensor readings, yields substantial performance gains.
Key Results
- MapSentinel achieved a median tracking accuracy of 1.96 m.
- It demonstrated a 31.3% improvement in tracking accuracy compared to a purely WiFi-based system.
- It also showed a 29.1% improvement over a fusion system combining WiFi and ultrasonic sensors without leveraging the floorplan's space use knowledge.
- The performance gains were most significant in constrained spaces (corridors) and static spaces (cubicles), where the contextual models are most effective.
- The "oracle" for context estimation achieved a mean accuracy of 52.41%. The paper notes that a perfect oracle would yield even better results, suggesting room for improvement in context prediction.
Limitations and Future Work
- "Oracle" Accuracy: The context estimation "oracle" is not perfect, with a mean accuracy of 52.41%. This noise in context estimation impacts the overall tracking accuracy.
- Multiple Occupant Tracking: The current system is designed for single-occupant tracking. Identifying and tracking multiple anonymous occupants simultaneously, especially when they are in close proximity or in the same detection zone, is a significant challenge.
- Integration with Building Systems: While the system demonstrates improved tracking, its integration into actual building control systems (lighting, ventilation) for energy efficiency is a future direction.
- Improving the "Oracle": The authors suggest that a more advanced "oracle" design could further enhance tracking performance.
Summary updated