UAVs using Bayesian Optimization to Locate WiFi Devices

Mattia Carpin, Stefano Rosati, Mohammad Emtiyaz Khan, Bixio Rimoldi

Introduction

In search-and-rescue (SAR) operations, time is the most critical variable. While most individuals carry WiFi-enabled devices, localizing these "non-collaborative" targets—devices not actively broadcasting their location—remains a significant challenge. Traditional search methods, such as aerial visual inspections, are often hampered by line-of-sight requirements and the heavy computational load of real-time image processing.

Using Received Signal Strength Indicators (RSSI) offers a viable alternative, but it introduces three primary hurdles: high stochasticity due to multipath fading and shadowing, data sparsity in large search regions, and a non-linear relationship between signal strength and distance in complex environments. Researchers at EPFL have addressed these issues by developing an active sensing system that utilizes fixed-wing Unmanned Aerial Vehicles (UAVs) and Bayesian Optimization to rapidly locate devices across large areas.

Methodology: Active Sensing via Bayesian Optimization

The core of this approach is treating signal strength mapping as a "black-box" optimization problem. Rather than attempting to model the complex physics of radio wave propagation, the system uses data-driven machine learning to map the environment dynamically. The process follows a three-step iterative loop:

  1. Gaussian Process (GP) Regression: The UAV utilizes collected RSSI measurements to construct a probabilistic model of signal strength across the search area. GPs are particularly effective as they provide both a prediction of signal strength and a quantified measure of uncertainty for unvisited locations.
  2. Acquisition Function: To determine the next flight destination, the system employs an "Expected Improvement" (EI) function. This function mathematically balances exploitation (visiting areas with known high signal strength) and exploration (investigating unknown areas with high uncertainty).
  3. Dynamic Trajectory: Unlike pre-planned grid searches, the UAV’s path is adaptive. After an initial baseline scan, it repeatedly navigates to locations that maximize the EI function.
Figure 1: The UAV's trajectory illustrating the balance between exploration and exploitation phases to localize the target WiFi device.

Because fixed-wing drones cannot hover, the navigation algorithms were adapted to fly in circles around the optimal points identified by the GP model, allowing for continuous data collection while maintaining flight speed.

Experimental Setup

The system was implemented on a SenseFly eBee, a fixed-wing UAV chosen for its high speed and energy efficiency. The aircraft carried an onboard Gumstix computer to execute the localization algorithms in real-time.

To evaluate the system, researchers conducted field tests in a $1000 \times 1000$ meter square field in Switzerland. A Samsung Galaxy Nexus served as the target device. Baseline data was also collected via exhaustive scans to characterize RSSI fluctuations across different hardware, such as laptops and smartphones, ensuring the model's robustness against varying signal profiles.

Results and Performance

The Bayesian Optimization approach demonstrated significant improvements over traditional exhaustive scanning methods. In real-world field experiments, the system successfully reduced the search area to a 100-meter radius around the target within just five minutes.

Figure 6: Localization error over time, showing convergence to less than 20 meters within a 15-minute flight window.

As shown in the performance data, the final localization error was reduced to less than 20 meters (specifically 18m) within a 15-minute flight. This represents a massive efficiency gain compared to exhaustive grid searches, which typically require over an hour to cover similar areas with lower precision.

Discussion and Future Work

While the current system is highly effective for stationary targets, several areas for future research remain. If a target is mobile, the Gaussian Process model would need to incorporate temporal decay or a dynamic state-space model to track movement. Additionally, while the system performed well in open fields, its efficacy in dense urban environments or heavy forests—where signal attenuation and multipath interference are more severe—requires further exploration. Future iterations may also look toward 3D localization to account for mountainous terrain or multi-story structures.

Summary updated

Abstract

We address the problem of localizing non-collaborative WiFi devices in a large region. Our main motive is to localize humans by localizing their WiFi devices, e.g. during search-and-rescue operations after a natural disaster. We use an active sensing approach that relies on Unmanned Aerial Vehicles (UAVs) to collect signal-strength measurements at informative locations. The problem is challenging since the measurement is received at arbitrary times and they are received only when the UAV is in close proximity to the device. For these reasons, it is extremely important to make prudent decision with very few measurements. We use the Bayesian optimization approach based on Gaussian process (GP) regression. This approach works well for our application since GPs give reliable predictions with very few measurements while Bayesian optimization makes a judicious trade-off between exploration and exploitation. In field experiments conducted over a region of 1000 $\times$ 1000 $m^2$, we show that our approach reduces the search area to less than 100 meters around the WiFi device within 5 minutes only. Overall, our approach localizes the device in less than 15 minutes with an error of less than 20 meters.
BibTeX
@misc{carpin2015uavs,
  author = {Mattia Carpin and Stefano Rosati and Mohammad Emtiyaz Khan and Bixio Rimoldi},
  title = {{UAVs using Bayesian Optimization to Locate WiFi Devices}},
  year = {2015},
  eprint = {1510.03592},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1510.03592},
}
Figure