Fusion-DHL

Fusion-DHL: WiFi, IMU, and Floorplans for Indoor Location Tracking

Indoor positioning has long challenged researchers seeking to replicate the seamless navigation we enjoy outdoors.

While GPS excels in open spaces, it fails inside buildings.

Fusion-DHL solves this by combining three data sources: WiFi signals, inertial measurement units (IMUs), and digital floorplans.

The Indoor Positioning Challenge

Accurately tracking indoor movements remains difficult. Wireless methods like WiFi or Bluetooth can help, but they rely on databases of signal information, drain battery, and can still be off by many meters.

Inertial navigation offers an alternative using smartphone IMU sensors. It reconstructs detailed trajectories without constant network communication. But it has a critical weakness: sensor errors accumulate over time, causing positions to drift further from reality with each step.

How Fusion-DHL Works

Fusion-DHL combines three data sources through a multi-stage process.

Stage One: Initial Trajectory Alignment

The system processes IMU data through RoNIN, a neural network that generates trajectories from motion sensors. It also receives sparse location updates from Google's Fused Location API (FLP), which provides WiFi-based position estimates once per minute.

These streams merge using non-linear least squares optimization—a technique that finds the best fit between the IMU trajectory and sparse WiFi anchor points. This places the trajectory accurately within the building.

Stage Two: Neural Network Refinement

The trajectory is divided into smaller segments using a sliding window. Each segment, paired with its corresponding floorplan patch, feeds into a convolutional neural network. The network analyzes the trajectory against the building's architecture and predicts displacement vectors for each point, nudging the path to align with hallways, rooms, etc.

Stage Three: Final Optimization

The system repeats the optimization and neural network steps with one key change: instead of using sparse FLP positions as constraints, it samples positions from the refined trajectory. This produces the final, highly accurate result.

Real-World Performance

Testing in a shopping mall demonstrates the system's capabilities. A subject walking for 44 minutes generated trajectories compared against several baselines:

• FLP trajectory with one location update per minute
• Pure inertial navigation from RoNIN aligned using only the first two FLP points
• A baseline method using conditional random field algorithms with the same input data

Fusion-DHL eliminated the accumulated errors of pure inertial navigation while aligning the trajectory to the floorplan's features. It reduced average WiFi positioning error from 12 meters to just 5 meters. More impressively, it generated 50 points per second compared to FLP's single point per minute.

The system also estimates body heading direction by applying angle corrections from the trajectory to the original RoNIN heading estimates.

Training with Synthetic Data

An innovative aspect of training involves synthetic data generation. Researchers created additional training examples through a simple process: hand-drawing paths on maps. This helped the network generalize across different building types and layouts.

Evaluation and Benchmarks

The research team created a benchmark dataset spanning over 15 hours of natural human motion across 42 kilometers. Ground truth came from sparse location points that subjects manually marked on floorplans in real-time during data collection.

Evaluation covered four datasets, categorized by:

• Sparsity of ground truth annotations
• Generalization capability across different building types

This rigorous testing confirmed the system's robustness across varied indoor environments.

Practical Applications

Dense location history tracking enables numerous applications:

Space Planning: Understanding how people move through buildings helps architects and facility managers optimize layouts for better flow and space utilization.

Transportation Optimization: Shopping malls, airports, and transit hubs can analyze movement patterns to reduce congestion and improve wayfinding.

Contact Tracing: The COVID-19 pandemic highlighted the importance of knowing who was where and when. Dense, accurate indoor positioning enables effective contact tracing while maintaining the privacy benefits of on-device processing.

Retail Analytics: Store owners can understand customer movement patterns to optimize product placement and store layout.

Technical Advantages

Fusion-DHL represents a significant leap forward in indoor navigation. The system produces trajectories that are:

• Orders of magnitude denser than current industrial solutions
• Twice as accurate as existing state-of-the-art systems
• Energy efficient, leveraging existing smartphone sensors
• Privacy-friendly through on-device processing

Combining non-linear least squares optimization with deep learning creates a hybrid approach that captures the strengths of both mathematical modeling and data-driven pattern recognition.

Looking Forward

As smartphones become more sophisticated and building information modeling becomes widespread, systems like Fusion-DHL will become essential infrastructure for indoor spaces where we spend most of our time. The technology bridges the gap between our outdoor GPS-enabled world and the complex indoor environments that have long resisted accurate positioning.

The future of indoor navigation lies not in choosing between WiFi, inertial sensors, or architectural data, but in fusing all available information. Fusion-DHL demonstrates that this integrated approach can finally deliver the accuracy and density needed for practical indoor positioning applications. However, the proposed system can only work offline, thus the real-time indoor localization remains an open problem.

Integrity note: This post reviews and references publicly available work from third-party authors.