TL;DR
This paper introduces a novel pedestrian navigation system that uses "invisible salient landmarks" – patterns in smartphone sensor data – to guide users along predesigned routes without relying on digital maps. By adapting the Dempster-Shafer theory of evidence, the system achieves high accuracy in both indoor and outdoor environments, outperforming traditional GPS and Wi-Fi methods.
What Problem Does This Paper Address?
Navigating unfamiliar environments, especially for "last mile" journeys (e.g., from a transit stop to a specific building), remains a challenge. While GPS and digital maps are common, they have limitations: GPS accuracy can be poor indoors or in urban canyons, and digital maps require constant attention. Existing sensor-based navigation often relies on a limited set of sensors or traditional localization techniques. This paper tackles the problem of providing robust, map-less pedestrian navigation by leveraging the rich, often underutilized, data from multiple smartphone sensors as a form of "invisible landmarks." These landmarks are not visual cues but rather dynamic patterns in sensor readings that can uniquely identify segments of a route.
Methods
The core of this approach is an enhanced Dempster-Shafer (D-S) theory of evidence. Here's a breakdown:
- Invisible Salient Landmarks: The system treats changes and patterns in smartphone sensor data (gyroscope, magnetometer, light, Wi-Fi, GPS) as "invisible landmarks." These are collected along a predesigned "business card route" (origin to destination).
- Route Segmentation: The collected sensor data is segmented based on significant changes in sensor readings. Each segment represents a distinct part of the route.
- Evidence Framework: For each route segment, a "frame of discernment" is created, capturing the combined characteristics of multiple sensors.
- Improved D-S Theory:
- Similarity Measure: A novel similarity measure is introduced to compare real-time sensor data with the predefined evidence framework.
- Sensor Weighting: A match-error-based approach is used to assign weights to sensors, giving more importance to those that provide more reliable and discriminative information for a given segment. This helps handle conflicts and improve accuracy.
- Conflict Handling: The improved D-S theory incorporates strategies to manage conflicts arising from sensor unreliability or environmental interference.
- Map-less Navigation: The system guides pedestrians by matching their current sensor readings to the pre-established evidence framework, allowing them to follow the route without needing to look at a map.
Datasets
The study used data collected from a specific experimental route on the campus of Wuhan University, from the university entrance to the No. 2 School Building. Data was collected 10 times using a VIVI X6 smartphone. The first collection was used to build the evidence framework, and the subsequent nine were used for training and calculating sensor weights. Nine labelled points (waypoints) were selected along the route to check matching errors and evaluate performance.
Key Insights and Ideas
- Sensor Fusion as Landmarks: The most significant insight is viewing the dynamic interplay of multiple smartphone sensor signals as "invisible landmarks." This moves beyond traditional landmark recognition (visual cues) to a more abstract, data-driven approach.
- Contextual Navigation: By segmenting routes based on sensor patterns, the system implicitly captures the environmental context of each segment, allowing for more robust navigation than simple coordinate-based methods.
- Adaptive Weighting: The match-error-based sensor weighting is crucial. It acknowledges that not all sensors are equally useful in all situations or segments, allowing the system to adapt and prioritize more reliable data sources.
- Map-less Viability: The success of this approach demonstrates that effective pedestrian navigation can be achieved without constant reliance on visual maps, which is particularly beneficial for users who prefer not to be distracted by screens.
Key Results
- High Match Success Rate: The proposed approach achieved a 100% match success rate for all nine labelled points across multiple data collections. This significantly outperformed the traditional D-S theory, which had an average success rate of about 40%.
- Improved Positioning Accuracy:
- Outdoors: The average positioning error was less than 5m, with a reduced percentage (RP) of positioning errors compared to GPS ranging from 15.9% to 54.4%.
- Indoors: The average positioning error was less than 3m, with an RP compared to Wi-Fi ranging from 10.1% to 62.6%.
- Robustness: The system demonstrated robustness against changes in speed and time elapsed between landmarks, treating intermediate sections as straight paths.
Limitations and Future Work
- Limitations:
- The study acknowledges that the performance can be affected by external interference (e.g., a parked vehicle affecting magnetometer readings).
- For certain points (like #6), positioning errors were higher, potentially due to environmental factors like dense trees blocking GPS or large intersection areas where sensor characteristics are less distinct.
- The current approach relies on predesigned routes.
- Future Work:
- Map Integration: Combine mapless navigation with digital maps to leverage both visual and invisible landmarks.
- Adaptive Sensor Weighting: Develop more adaptive methods for assigning sensor weights, potentially smartphone-dependent.
- Invisible Landmark Maps: Create navigation maps based on these invisible landmarks and integrate them into existing point-of-interest (POI) systems.
- Comparative Navigation: Explore comparative navigation strategies in future work.
Summary updated