💻 computer science

Reflection-Based Relative Localization for Cooperative UAV Teams Using Active Markers

This paper introduces a novel reflection-based method that leverages environmental reflections of active markers to achieve robust, long-range relative localization for heterogeneous UAV teams without requiring prior knowledge of robot size or surface properties, while explicitly accounting for surface irregularities like dynamic water.

Original authors: Tim Lakemann, Daniel Bonilla Licea, Viktor Walter, Martin Saska

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tim Lakemann, Daniel Bonilla Licea, Viktor Walter, Martin Saska

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a friend in a dark, crowded room, but you can't talk to them, and you don't know how tall they are or what they are wearing. You only have a flashlight. Usually, you would shine your light directly at them to see where they are. But what if, instead of looking at your friend directly, you looked at the reflection of their flashlight on the shiny floor?

That is essentially what this paper does, but with drones (UAVs) instead of people in a room.

The Problem: The "Ghost" in the Machine

When drones fly together, they need to know exactly where their teammates are to avoid crashing. Usually, they use special lights (markers) that blink in unique patterns. However, there's a catch:

Direct Line of Sight: If the drones get too far apart, the blinking lights look too small to see clearly.
The "Ghost" Problem: In the past, engineers treated reflections on the floor or water as annoying "ghosts" or noise that confused the cameras. They tried to filter them out using special glasses (polarization filters) or complex lighting.

The Solution: Turning a Bug into a Feature

This research team decided to stop fighting the reflections and start using them. They realized that the reflection of a drone's light on a surface (like a polished warehouse floor or a calm lake) actually contains a lot of useful information.

Think of it like this: If you shine a laser pointer at a mirror, the dot you see on the wall tells you exactly where the mirror is. If you shine it at a wavy pond, the reflection stretches out. The team figured out that by measuring how much that reflection stretches and where it sits, they can calculate exactly where the other drone is, even if they can't see the drone itself clearly.

How It Works (The "Flashlight and Floor" Trick)

Here is the step-by-step process the paper describes, simplified:

The Blinking Code: Each drone has a set of UV lights (invisible to the human eye but visible to the camera) that blink in a secret code. This ensures Drone A knows which light belongs to Drone B.
The Reflection Hunt: Drone A looks down at the ground. It sees two things:
- The direct light from Drone B (if visible).
- The reflection of Drone B's light bouncing off the floor or water.
The "Stretch" Factor: Because the ground isn't perfectly smooth (it might be rippled water or a slightly bumpy floor), the reflection doesn't look like a perfect dot. It looks like a stretched-out oval or an ellipse.
The Math Magic: The drone's computer builds a 3D "cone" of possibilities.
- One cone represents where the direct light could be.
- A second, wider cone represents where the reflection could be, accounting for the wobbly water or bumpy floor.
- Where these two cones overlap is the most likely spot where the other drone is flying.

Why This Is a Big Deal

The paper highlights three main superpowers of this new method:

It Doesn't Care About Size: You don't need to know if the other drone is a tiny insect-sized robot or a large delivery drone. The math works either way because it relies on the light's reflection, not the size of the drone.
It Works in the Dark and on Water: Most camera systems struggle in the dark or on moving water. This method actually likes the water. Even if the waves make the reflection wobble, the system accounts for that "wobble" and still finds the drone.
It Sees Further: The paper tested this against the current best method (called UVDAR). The old method struggled to see drones past 20 meters. This new reflection method worked reliably at 30 meters and beyond.

The Experiments

The team tested this in two main ways:

Indoors: They flew drones over shiny office floors (PVC) and tiled hallways, both in bright daylight and in the dark. The system worked perfectly, finding the drones with high accuracy.
Outdoors: They flew over a lake. The wind made the water ripple, creating a messy, dancing reflection. Even with the water moving, the system successfully tracked the other drone.

The Bottom Line

This paper introduces a clever way for drones to find each other by using the floor as a mirror. Instead of trying to clean up the "messy" reflections on the ground, they turned that mess into a powerful tool. It allows drone teams to fly closer together, further away, and in more difficult environments (like over water or in the dark) without needing expensive extra sensors or knowing exactly what their teammates look like.

Technical Summary: Reflection-Based Relative Localization for Cooperative UAV Teams

Problem Statement
Reliable relative localization is critical for the safe execution of collaborative tasks by multi-robot systems. While Global Navigation Satellite Systems (GNSS) are standard for outdoor operations, their reliability degrades significantly in dense forests, urban canyons, and natural landscapes, and they are entirely unavailable indoors. Existing alternatives present specific limitations: Real-Time Kinematic (RTK) systems constrain operations to the range of a ground station; motion capture systems lack the flexibility for unknown environments; and vision-based methods using passive markers (e.g., AprilTags) impose constraints on system dynamics and marker size. Furthermore, active marker systems (e.g., UVDAR) typically require prior knowledge of the marker configuration and suffer from reduced accuracy as distance increases, particularly for Micro Aerial Vehicles (MAVs) where the physical spacing between markers becomes insufficient for reliable pixel separation. Additionally, surface reflections from active markers are traditionally viewed as a source of ambiguity and noise, often mitigated using polarization filters or structured light, rather than being utilized for localization.

Methodology
The proposed approach introduces a novel reflection-based relative localization system that exploits surface reflections of active markers attached to team members. The system operates without prior knowledge of robot size, marker configuration, or surface properties.

Theoretical Foundation: The method models the interaction between unpolarized light (emitted by UAV LEDs) and reflective surfaces. It accounts for both specular and diffuse reflections, acknowledging that surface irregularities (e.g., water waves, textured floors) cause the reflection to spread. The system derives the reflection power coefficient based on Fresnel equations and models surface roughness using Gaussian height distributions or capillary-gravity wave models for dynamic water surfaces.
Detection and Clustering: Each UAV emits a unique blinking sequence via UV-LEDs. The observer UAV detects these signals and their reflections. To handle high-exposure settings required for reflection detection, where a single LED may appear as multiple bright dots, the system employs a clustering algorithm. Detections within a specific radius are grouped; the dominant cluster represents the direct light source ( $p_d$ ), while a secondary cluster located vertically below is identified as the reflection candidate ( $p_r$ ). If the blinking sequence cannot be extracted from the reflection, $p_r$ is identified geometrically based on vertical alignment relative to $p_d$ .
Geometric Modeling: The system constructs two elliptical cones to represent uncertainty:
1. Direct Source Cone: Represents the uncertainty around the direct bearing vector ( $\vec{b}_d$ ) due to camera resolution and detection inaccuracies.
2. Reflection Cone: Represents the uncertainty around the reflection vector ( $\vec{b}_r$ ). This cone is constructed based on the spread of the reflection on the surface, defined by the major and minor axes of an ellipse centered at the reflection centroid.
Position Estimation: The intersection of these two cones defines the possible location of the transmitting UAV. Since the intersection of two cones results in a complex non-linear system, the authors employ a random sampling-based approach. Vectors are sampled from the reflection cone and tested against the direct source cone constraints. Valid samples (inliers) are fed into a particle filter with a constant-velocity motion model and a robust Cauchy kernel for weight updates. The final position estimate is derived from the weighted particle cloud.
Constraints: The method assumes operation over ground or water surfaces where a single dominant reflection can be reliably obtained. It does not rely on wall reflections, as urban walls typically produce diffuse reflections with insufficient intensity and require knowledge of wall orientation.

Key Contributions

Theoretical Derivation: The work provides the theoretical background and physics underlying reflection-based localization, modeling the interaction of light with various surface properties.
Novel Algorithm: A camera-only relative localization approach is proposed that requires only the range to the reflective surface (e.g., from a downward-facing rangefinder or barometer) and no prior knowledge of team member sizes or surface properties.
Experimental Validation: The approach is validated across heterogeneous platforms (large UAVs and small MAVs) in both indoor and outdoor environments, demonstrating robustness to varying lighting conditions and surface types (PVC, tiles, water).

Results

Indoor Performance: Experiments conducted on PVC and tiled floors under daylight and nighttime conditions showed reliable operation. The Mean Absolute Error (MAE) was consistently lowest along the vertical (z) axis (approx. 0.08–0.12 m) and higher along the optical (x) axis (approx. 1.22–2.12 m), which is expected for monocular depth estimation. The system successfully ran in real-time on low-computational-power hardware (Raspberry Pi 5).
Outdoor Performance: In outdoor experiments over water, the method achieved an effective range exceeding 30 meters.
Comparison with State-of-the-Art: When compared to the UVDAR method (which relies on direct line-of-sight detection of known marker spacing):
- The reflection-based method maintained high accuracy at distances beyond 30 m, whereas UVDAR struggled beyond 20 m.
- Along the optical axis (x), the proposed method reduced MAE by 18.5% to 37.3% compared to UVDAR, with statistically significant improvements (p-value < 0.05).
- Along the lateral axis (y), improvements of 13.9% to 58.7% were observed in specific experiments.
- Vertical (z) accuracy was comparable to UVDAR, with no statistically significant difference.
Robustness: Sensitivity analysis indicated that systematic height biases of up to ±15% resulted in only minor variations in localization accuracy, confirming the method's robustness.

Significance
The paper claims that this work represents the first exploration of relative localization for multi-robot systems using surface reflections. By transforming a typically unwanted artifact (reflections) into a useful signal, the method extends the operational range of vision-based localization and removes the dependency on predefined marker configurations. This enables efficient, compact solutions for close cooperative formation flights in confined indoor environments and marine operations, particularly where GNSS is unavailable and traditional marker-based systems are limited by size or lighting constraints. The approach offers a lightweight, cost-effective alternative that enhances the autonomy of heterogeneous UAV teams.