[Indoor Positioning] Eliminating the Site Survey: Automatic Radio Map Construction via Crowdsourcing

TL;DR

Building a WiFi radio map for indoor positioning usually requires a PhD-level amount of manual labor. This paper introduces a system that automates the process by harvesting "outdoor-to-indoor" transitions from normal users. By using the phone's light sensor to identify building entrances and optimization algorithms to "stitch" PDR traces together based on WiFi similarity, the authors achieve sub-3-meter accuracy without ever seeing a floor plan.

The "Mapping" Bottleneck

Indoor positioning based on WiFi Fingerprinting is ubiquitous (think Google/Apple Maps), but it has a dark secret: it traditionally requires "War-driving" or manual site surveys. Professionals must walk through every corridor, recording WiFi RSSI and labeling ground truth coordinates.

While modern researchers have tried Crowdsourcing, they usually hit three walls:

1. PDR Drift: Gyroscopes and accelerometers accumulate error over time, distorting the trace shape.
2. Point of Origin: Without a floor plan or GPS (which fails indoors), we don't know where a trace actually starts in the real world.
3. Rotation Bias: Distorted magnetic fields indoors make the "North" of a phone's compass unreliable.

Methodology: Anchor, Correct, and Merge

The authors' core insight is to treat users entering/exiting a building as "Anchors."

1. The Gateway Anchor (Outdoor-Indoor Processing)

When a user walks from outside to inside, the phone's light sensor sees a massive drop in intensity. The system marks this as a "Gate Point." By correlating the last good GPS coordinate before the gate and the first one after the exit, the system performs a Procrustes-style transformation (translation, rotation, and scaling) to align the indoor segment with Global Geographic coordinates (GEOREF).

2. Solving the Loop Closure (Indoor Processing)

For traces purely within the building, the authors use a "Force-Directed" intuition. If a user walks in a loop and sees the same WiFi environment twice, those two PDR points should be at the same physical location.

The system partitions the trace into segments at every turn > 60°. It then solves an optimization problem: How much should we rotate each segment to minimize the "Force" (physical distance) between points with nearly identical WiFi signatures?

Fig. 1: Evolution of traces from raw, drifted data to a coherent road path.

3. Final Integration

Once a "Skeleton" of road paths is formed by the outdoor-indoor traces, pure indoor traces are "clicked" into place by matching their WiFi fingerprints to the anchored skeleton.

Experimental Results

The system was tested in a large shopping mall (3600m²).

• Road Path Accuracy: The generated map turned out to be remarkably accurate, with a median error of 2.8m.
• Positioning Performance: Using this "zero-effort" map, the system achieved a 2.9m median positioning accuracy.

Fig. 2: Cumulative Distribution Function (CDF) of positioning errors. Note how WiFi-recalibration prevents the linear error growth seen in pure PDR.

Critical Insight: The Power of Context

The brilliance of this work lies in its context-awareness. Instead of trying to fix PDR sensors through raw signal processing—a losing battle—it uses environmental context (light sensors, GPS transitions, and the "closure" of walking paths) to constrain the mathematical solution.

Limitations & Future Work

• Uniformity: The current model assumes a horizontal phone carry, which doesn't account for users with phones in pockets or swinging hands.
• Scalability: While 3600m² is a solid test, mega-malls with complex multi-floor architectures would require 3D road path merging.

Conclusion

This paper paves the way for truly scalable smart city services. By removing the need for floor plans and manual surveys, it moves us closer to a world where any building can be "mapped" simply by its visitors walking through it.