💻 computer science

PATH: Continuous Target Sensing among Autonomous Cooperative Drones

This paper presents PATH, a lightweight, geometry-assisted framework that enables autonomous UAVs to seamlessly transfer target tracking responsibilities by reconstructing targets in 3D, projecting them into the receiver's view as spatial priors, and verifying acquisition through a mutual agreement handshake, thereby achieving high accuracy and low latency despite visual ambiguities and limited communication bandwidth.

Original authors: Heegyeong Kim, Alice James, Avishkar Seth, Endrowednes Kuantama, Jane Williamson, Yimeng Feng, Richard Han

Published 2026-09-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Heegyeong Kim, Alice James, Avishkar Seth, Endrowednes Kuantama, Jane Williamson, Yimeng Feng, Richard Han

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 a world where small flying robots, known as drones, work together to watch over wildlife, monitor construction sites, or search for people in distress. These machines are incredibly useful, but they share a common limitation: their batteries run out. A single drone cannot stay in the air forever. To solve this, engineers have long dreamed of a system where one drone could pass the job of watching a specific target to a fresh drone, allowing the first to land and recharge while the mission continues without interruption. The challenge, however, is not just finding a new drone; it is ensuring that the new drone is looking at the exact same person, animal, or object that the first drone was tracking. Because the two drones are moving, they see the world from different angles and distances. A person might look small and distant to one drone but large and close to the other, or they might be surrounded by other people who look very similar. If the second drone grabs the wrong target, the mission fails.

Researchers at Macquarie University have developed a new method called PATH to solve this tricky problem of passing the baton between flying robots. Instead of relying on global satellite positioning, which can be unreliable in cities or forests, or trying to match the visual appearance of the target like a human recognizing a face, this system uses simple geometry and a bit of math to ensure the two drones agree on what they are seeing. The process begins when the first drone, the sender, identifies a target and measures its distance and position in three-dimensional space. It then sends this location data to the second drone, the receiver. The receiver, which can see a special marker on the first drone to know exactly where it is relative to the sender, uses that information to calculate where the target should appear in its own camera view. It essentially draws a map in its mind of where to look.

Once the second drone spots a potential target near that predicted spot, it does not just assume it is the right one. Instead, it sends a description of what it found back to the first drone. The first drone then checks if this new description matches what it is currently seeing. This check is called a "Mutual Agreement Handshake." The system waits until both drones have agreed on the target for a specific period of time—about seven seconds of continuous confirmation—before officially transferring the tracking duty. This ensures that the handoff is not a mistake caused by a momentary glitch or a confusing visual trick.

The researchers tested this system with real drones flying indoors and outdoors, tracking everything from small robot cars to walking humans. In one set of tests, they created a difficult scenario where several people wore identical clothing, making it nearly impossible for the drones to tell them apart just by looking at their faces or clothes. In these confusing situations, the new method was remarkably successful. It correctly identified the intended target 96 percent of the time, while older methods that relied on matching visual features often got confused and picked the wrong person. The system also proved to be very efficient, requiring very little data to be sent between the drones and running smoothly on small, low-power computers without slowing them down.

The study also looked closely at what could go wrong. The researchers found that the most significant source of error was not the distance measurement, but the uncertainty in knowing exactly how the two drones were oriented relative to each other. However, even with these small errors, the system was robust enough to handle the task. The method works without needing to attach any special equipment to the target being watched, making it suitable for delicate situations like observing wildlife without disturbing them. By focusing on the geometric relationship between the drones and the target, rather than trying to recognize the target's appearance, the researchers have created a reliable way for flying robots to cooperate. This approach offers a practical solution for extending the life of drone missions, ensuring that when one robot gets tired, another can seamlessly take over the watch without losing sight of the goal.

Technical Summary: PATH (Perspective Alignment & Tracking Handoff)

Problem Statement
Continuous target sensing by Uncrewed Aerial Vehicles (UAVs) is often constrained by limited flight endurance, necessitating cooperative operations where tracking responsibility is transferred from a sender UAV (D1) to a receiver UAV (D2). The core challenge lies in the "handoff": D2 must identify and acquire the exact same physical target currently tracked by D1 from a different, time-varying viewpoint. Existing solutions face significant limitations:

Global Localization: Relying on global GNSS coordinates is often insufficient for distinguishing closely spaced targets due to multipath errors, signal blockage, and the need for infrastructure (e.g., RTK). Furthermore, global position does not directly resolve which of several nearby observations corresponds to the sender's target.
Active Indication: Laser-based designation requires additional active pointing mechanisms and can disturb wildlife or interfere with optical sensing, limiting its utility in passive observation scenarios.
Appearance-Based Association: Feature matching or learned re-identification struggles under significant viewpoint and scale changes, weak visual features, or the presence of visually similar distractors. These methods also impose heavy computational loads on onboard hardware.

Methodology: The PATH Framework
The paper presents PATH (Perspective Alignment & Tracking Handoff), a platform-agnostic, geometry-assisted framework designed to verify target handoffs without instrumenting the target or relying on global positioning. The system operates through a five-step procedure:

Local Coordinate Frame Establishment: D2 observes a fiducial marker rigidly mounted on D1 to estimate the relative pose ( $^vT_2$ ) between the two UAVs within a handoff-local coordinate frame ( $F_v$ ) anchored to D1.
3D Target Reconstruction: D1 reconstructs the tracked target as a metric 3D point ( $^vX_t$ ) using onboard RGB-D sensing (Intel RealSense D415). This involves undistorting the pixel coordinates, applying the depth measurement, and transforming the point from the D1 camera frame to the local frame $F_v$ .
Spatial Prior Projection: D1 transmits the 3D target point to D2. D2 transforms this point into its own camera frame and projects it into its image plane, creating a spatial prior (a predicted pixel location) for target acquisition.
Candidate Acquisition: D2 runs its local detector around the predicted location to generate a candidate bounding box ( $B^c_2$ ). Crucially, D2 does not transmit raw images or feature descriptors; it only returns the bounding box coordinates.
Mutual Agreement Handshake: D2 sends the candidate bounding box back to D1. D1 transforms this box into its own image frame, adjusting dimensions based on relative depth and focal length. D1 then calculates the Intersection over Union (IoU) between the transformed candidate box and its own active target box.
- Verification: A handoff is only confirmed if the IoU exceeds an adaptive threshold (scaled by apparent target size) for $N$ consecutive frames (e.g., 350 frames at 50 Hz). This temporal consistency prevents transient errors from triggering incorrect handoffs.

Key Contributions

Formulation: The authors formulate cooperative UAV handoff as a cross-view sensing problem requiring verification of the same physical target before responsibility transfer.
Geometric Framework: Development of PATH, which combines RGB-D target reconstruction, fiducial-based relative pose estimation, and cross-view projection to transfer target locations. It replaces appearance matching with geometric consistency verification.
Implementation & Evaluation: Comprehensive evaluation on heterogeneous UAV platforms (DJI Tello and custom Pixhawk/Jetson Nano quadrotors) covering geometric accuracy, target acquisition under visual ambiguity, sensor sensitivity, and onboard resource usage.

Experimental Results

Geometric Accuracy: In controlled static indoor experiments (420 measurements), PATH achieved a mean relative-position error of 0.047 m and a mean target-position error of 0.030 m.
Target Acquisition Accuracy: In outdoor experiments with visually ambiguous conditions (multiple humans in similar clothing), PATH achieved a 96.0% frame-level receiver-side target acquisition accuracy. This resulted in 2.0% false-positive and 2.0% false-negative rates.
- Comparison: This significantly outperformed appearance-based baselines: ORB (62.1% false-negative rate) and XFeat (15.4% false-positive rate).
Sensor Sensitivity: Analysis revealed that relative-pose uncertainty (from the fiducial marker) is the dominant contributor to projection error, rather than target depth errors.
System Performance: The system operates at video rate (30–60 Hz) on resource-constrained hardware (Jetson Nano) using CPU-only processing for the handoff logic. Inter-UAV communication is extremely lightweight (<16 kB/s), transmitting only compact geometric data (3D points and bounding boxes) rather than video streams or feature descriptors.
Handoff Latency: Successful handoffs (42 out of 45 attempts) showed a median time-to-commit of 7.20 seconds after the IoU threshold was first met, with low timing jitter.

Significance and Claims
The paper claims that PATH provides a verified, lightweight, and robust mechanism for transferring target tracking between UAVs. Its primary significance lies in:

Decoupling from Global Infrastructure: It avoids reliance on high-precision GNSS or active laser designation, making it suitable for environments where global positioning is degraded or where passive observation is required.
Robustness to Ambiguity: By using geometric priors to constrain the search space, it effectively handles visual ambiguity where appearance-based methods fail.
Resource Efficiency: It demonstrates that accurate handoff verification can be achieved with minimal computational and communication overhead, preserving onboard resources for detection and control tasks.
Verification Layer: It acts as a complementary verification layer to existing detectors, ensuring that the receiver is tracking the correct target before the handoff is finalized, rather than replacing the detection algorithms themselves.

The authors note limitations, including dependence on the visibility of the fiducial marker and sensitivity to relative-pose errors, and suggest future work in markerless pose estimation and handling occlusions.

Technical Summary: PATH (Perspective Alignment & Tracking Handoff)

More like this