PaperlayerView exact revision on arXiv

Source-linked AI summary

PATH: Continuous Target Sensing among Autonomous Cooperative Drones

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

arXiv:2609.12456v1cs.RO

TL;DR

Limited UAV endurance motivates transferring tracking between platforms, but the receiver must identify the sender’s same target across differing viewpoints and visually similar candidates. PATH uses local geometry to guide acquisition and verifies the receiver’s candidate before transfer; experiments report low positioning errors and high acquisition accuracy in visually ambiguous scenes.

  • Problem

    Limited UAV endurance motivates cooperative tracking, but the receiver must identify the same physical target as the sender across moving viewpoints and visually similar candidates.

    Source: §I · p. 1

  • Method

    PATH reconstructs the sender’s target as a local 3D point, projects it into the receiver image as a spatial prior, and verifies the returned candidate through a Mutual Agreement Handshake.

    Source: §I · p. 2

  • Results

    Indoor geometric evaluation over 420 measurements yielded mean relative UAV position and target position errors of 0.047 m and 0.030 m, while ambiguous outdoor acquisition reached 96.0% accuracy with 2.0% false-positive and 2.0% false-negative rates.

    Source: §VII · pp. 9–10

  • Takeaways & Limitations

    The results support geometry-assisted handoff as an accurate, lightweight, platform-agnostic verification layer for cooperative UAV tracking.

    Source: §VII · p. 10

    The current evaluation covers two-UAV handoff under tested indoor and outdoor conditions, and relative-pose and depth errors affect projection accuracy.

    Source: §VI · p. 9

Abstract

from arXiv · show

Continuous target sensing by uncrewed aerial vehicles (UAVs) is constrained by limited flight endurance, motivating the transfer of tracking responsibility between cooperating UAVs. Such a handoff requires the receiver to identify the same physical target currently tracked by the sender despite differences in viewpoint, scale, and target appearance. Existing approaches based on global target localization or appearance-based cross-view association are limited by positioning uncertainty or ambiguous visual features. This paper presents Perspective Alignment \& Tracking Handoff (\textbf{PATH}), a platform-agnostic, geometry-assisted sensing and verification framework for target handoff between two moving UAVs. The sender reconstructs the tracked target as a metric 3D point using RGB-D sensing, while the receiver estimates its relative pose from a fiducial observation and projects the transmitted target point into its own image as a spatial prior for target acquisition. The receiver-generated candidate is then returned to the sender and verified through a cross-view Mutual Agreement Handshake before tracking responsibility is transferred. Real-world UAV experiments show mean relative-position and target-position errors of 0.047~m and 0.030~m, respectively. Under visually ambiguous conditions, PATH achieves 96.0\% frame-level receiver-side target acquisition accuracy, with 2.0\% false-positive and 2.0\% false-negative rates. A sensor-error sensitivity analysis shows that relative-pose uncertainty is the dominant contributor to receiver-view projection error. The implementation operates at video rate with compact inter-UAV communication below 16~kB/s at 60~Hz, demonstrating the feasibility of lightweight geometry-assisted target handoff on resource-constrained UAV platforms.

I. INTRODUCTION

Long-duration UAV tracking motivates handing responsibility to a second platform, but the receiver must identify the same physical target across moving viewpoints and visually similar candidates. PATH uses local geometry to guide receiver-side acquisition and verifies the returned candidate before transfer.

I. INTRODUCTION

  • Limited endurance motivates cooperative UAV handoff, where the receiver must acquire the same physical target tracked by the sender.Independent detection on the receiver can identify candidates but cannot by itself establish which corresponds to the sender’s target.Source: §I · p. 1
  • Changing viewpoints and visually similar objects make cross-view correspondence a distinct challenge beyond detecting an object in the receiver view.The handoff must establish correspondence between observations from two moving viewpoints.Source: §I · pp. 1–2
  • PATH reconstructs the sender’s target as a local 3D point, projects it into the receiver image as a spatial prior, and verifies the returned candidate through a Mutual Agreement Handshake.The target remains uninstrumented, and tracking responsibility transfers only after geometric verification.Source: §I · p. 2
  • The framework combines RGB-D target reconstruction, relative-pose sensing, cross-view projection, and candidate verification, and is evaluated on heterogeneous UAV platforms.The evaluation includes geometric accuracy, ambiguous-scene acquisition, handoff reliability, sensor-error sensitivity, and onboard costs.Source: §I · p. 2

A. Target Localization and Active Target Indication

The related work covers global and active target indication, appearance-based association, and cooperative localization. PATH instead constructs local handoff geometry and uses it to constrain receiver acquisition and verify the sender’s target.

A. Target Localization and Active Target Indication

  • GNSS-based target localization may not distinguish closely spaced targets under degraded positioning, while laser indication adds pointing requirements and can disturb wildlife.PATH instead uses passive onboard visual sensing and leaves the tracked target uninstrumented.Source: A. Target Localization and Active… · p. 2

B. Cross-View Visual Association

  • Appearance matching can struggle with similar targets and major viewpoint or scale changes; PATH constrains the expected receiver-image location using camera geometry instead.The receiver detector generates a candidate for verification rather than matching target appearance across views.Source: B. Cross-View Visual Association · pp. 2–3

C. Cooperative Relative Localization and Geometric Sensing

  • Cooperative localization primarily estimates relative or shared states, whereas PATH addresses whether a receiver detection is the target tracked by another moving UAV.A fiducial observation supplies relative pose, RGB-D sensing reconstructs the target as a metric 3D point, and geometric agreement verifies the returned candidate.Source: C. Cooperative Relative Localization… · p. 3

III. PROPOSED PATH VERIFICATION SYSTEM

  • PATH’s five-step handoff establishes a local frame, transfers the sender’s target location, acquires and returns a receiver candidate, then checks bounding-box IoU.The IoU evaluation completes handoff verification.Source: §III · p. 3

A. Design Objectives and System Choices

  • PATH is a detector- and tracker-complementing verification layer that requires bounding boxes, not a particular target category or feature-level appearance matching.It uses a marker mounted on the sender and compact inter-UAV messages rather than raw image or video transmission.Source: A. Design Objectives and System Choices · p. 3

B. Handoff Geometric Representation

PATH expresses the sender’s target and receiver pose in a local frame anchored to the sender, then projects the target point into the receiver image as a spatial prior. The receiver’s independently generated bounding box is returned for verification.

B. Handoff Geometric Representation

  • The handoff-local frame is anchored at the center of the fiducial marker on D1, giving the target observation and D2 pose a shared spatial reference.The coordinate axes follow the orientation shown in Fig. 2.Source: B. Handoff Geometric Representation · p. 3, Figure 2 · p. 3
  • D2 observes D1’s fiducial marker to estimate its pose relative to the handoff-local frame, including its position and orientation.That relative pose is used to transform the target point into D2’s camera frame.Source: B. Handoff Geometric Representation · p. 3
  • D1 uses the tracked bounding-box center and registered depth to reconstruct the target in its camera frame, then expresses that point in the local frame.The resulting 3D target point is transmitted to D2.Source: B. Handoff Geometric Representation · pp. 3–4
  • D2 projects the transformed target point into its image as a spatial prior, then generates its candidate bounding box independently for return to D1.The returned detector-generated candidate is used in the Mutual Agreement Handshake.Source: B. Handoff Geometric Representation · p. 4

C. Mutual Agreement Handshake

The Mutual Agreement Handshake verifies that the receiver’s independently detected candidate agrees geometrically with the target tracked by the sender before responsibility transfers. It compares observations in a common image frame and requires scale-adjusted agreement to persist over time.

C. Mutual Agreement Handshake

  • PATH verifies the receiver’s independently detected candidate against the sender’s tracked target through geometric agreement, rather than feature-level appearance matching.The handshake occurs before tracking responsibility transfers.Source: C. Mutual Agreement Handshake · p. 4
  • The receiver-generated box is returned to the sender and mapped into the sender image, with its dimensions adjusted using relative image scale.The width and height approximation assumes locally constant target depth across the bounding-box extent; identical camera intrinsics make focal-length ratios unity.Source: C. Mutual Agreement Handshake · p. 4
  • The verification uses complete bounding-box overlap to capture positional and spatial-extent agreement, with an adaptive IoU threshold accounting for apparent target scale.Expected target area follows an inverse-square relationship with viewing distance, and the threshold is progressively relaxed as apparent size decreases.Source: C. Mutual Agreement Handshake · pp. 4–5
  • A handoff requires agreement to persist for N consecutive frames; a below-threshold overlap or missing candidate resets the sequence.The temporal criterion guards against transient detections and sensing errors.Source: C. Mutual Agreement Handshake · pp. 4–5

D. Sensor Sensitivity

The sensitivity analysis propagates target-depth and fiducial-pose perturbations through a nominal receiver-view projection. At 1.0 m, relative-pose perturbations contribute more RMS projection sensitivity than target-depth perturbations.

D. Sensor Sensitivity

  • The analysis propagates D415 depth-error and fiducial translation and rotation reference magnitudes through a nominal projection, characterizing geometric sensitivity rather than a deployed-system accuracy bound.The cited marker measurements differ from PATH's marker size and operating distance, so they are reference magnitudes, not PATH error bounds.Source: D. Sensor Sensitivity · p. 5
  • The nominal sensitivity geometry uses a rectified, zero-distortion pinhole model, while deployed PATH projection retains calibrated lens distortion; no separate intrinsic-calibration perturbation is included.The deployed unit lacks an available intrinsic-parameter covariance or reprojection-error distribution for such a perturbation.Source: D. Sensor Sensitivity · p. 5
  • At 1.0 m, combined RMS projection sensitivity is 17.41 pixels; it rises from 12.75 to 21.57 pixels across the evaluated 0.6–1.5 m depth range.The combined value is approximately 0.79% of the 1920 × 1080 receiver image diagonal.Source: D. Sensor Sensitivity · p. 6
  • At 1.0 m, fiducial-marker relative-pose sensitivity is 17.22 pixels, versus 2.56 pixels for target-depth sensitivity, making relative pose the dominant contributor under the nominal geometry.The analysis motivates using the projected point as a spatial search prior, with region-based Adaptive-IoU verification tolerating moderate projection and detection variation.Source: D. Sensor Sensitivity · p. 6

IV. EXPERIMENTAL SETUP

The evaluation used heterogeneous UAV platforms, individually calibrated cameras, and uninstrumented targets. Tests covered indoor and outdoor conditions and included measurements of geometry, handoff, and system costs.

IV. EXPERIMENTAL SETUP

  • PATH was evaluated on DJI Tello pairs and custom 425 mm quadrotors with Pixhawk Cube Orange controllers and Jetson Nano modules; quantitative geometric and sensitivity evaluations used the D415-equipped custom platform.The Tello pair tested PATH on different UAV hardware, rather than serving as the platform for the quantitative geometric and sensitivity analyses.Source: §IV · p. 6
  • Both platforms' cameras were individually checkerboard-calibrated, custom UAVs carried D415 RGB-D cameras mounted at approximately 45°, and the sender carried a 40 × 40 mm fiducial marker.Tracked targets were not instrumented.Source: §IV · p. 6
  • Experiments used a robot car, micro-UAVs, and a walking human in indoor controlled-lighting, no-wind conditions and outdoors with natural illumination and background clutter.Indoor tests additionally used ultrasonic positioning to assess estimated UAV and target positions.Source: §IV · p. 6
  • The evaluation covered receiver-view projection, cross-view verification, appearance-matching baselines, runtime, and communication; handoff acceptance required agreement for 350 consecutive frames at 50 Hz, a 7 s window.Inter-UAV communication used Wi-Fi and compact geometric observations; six volunteers participated in data collection.Source: §IV · p. 6

A. Receiver-View Projection Consistency

Static positioning tests quantified PATH's relative and target-position errors across multiple receiver configurations. A separate dynamic indoor experiment examined whether transferred spatial information remained consistent during motion.

A. Receiver-View Projection Consistency

  • The static evaluation used seven D2 positions and orientations, three target locations, and 60 measurements per D2 case, totaling 420 measurements against manually measured reference geometry.The setup is illustrated in Fig. 5a; Fig. 5b compares estimated D2 position across relative orientations.Source: A. Receiver-View Projection Consistency · p. 6, Figure 5 · p. 7
  • Mean static positional error was 0.047 m for D2 relative to D1 and 0.030 m for target position.The relative-position error was 0.038 m when D2 faced the same direction as D1 and 0.054 m when D2 faced toward D1 with relative yaw.Source: A. Receiver-View Projection Consistency · pp. 6–7, Figure 5 · p. 7
  • The dynamic indoor experiment showed transferred spatial information remaining consistent during motion, while D2 continuously observed D1's fiducial marker and used the transferred target location to identify the target in its own view.Together, the dynamic and static experiments examine complementary aspects of the handoff-local representation.Source: A. Receiver-View Projection Consistency · pp. 6–7, Figure 4 · p. 6

B. Receiver-Side Target Acquisition under Visual Ambiguity

The receiver must identify the sender’s target among visually similar candidates despite viewpoint and appearance variation. PATH uses the transferred geometric prior to focus acquisition, while feature-matching baselines can confuse distractors or fail to find correspondence.

B. Receiver-Side Target Acquisition under Visual Ambiguity

  • 96.0% frame-level target-acquisition accuracy, with 2.0% false-positive and 2.0% false-negative rates, was achieved by PATH in visually ambiguous outdoor experiments.The evaluated outcomes included 350 frames: 317 positive and 33 negative cases.Source: Figure 6 · p. 7, B. Receiver-Side Target Acquisition… · p. 7
  • The 350-frame evaluation classified outcomes as correct acquisition, selection of an incorrect candidate, or failure to acquire a valid target.It included six individuals wearing similar clothing and caps to create visually similar receiver-side candidates.Source: B. Receiver-Side Target Acquisition… · p. 7
  • 62.1% ORB false-negative and 15.4% XFeat false-positive rates contrasted with PATH’s 2.0% rates under visually similar outdoor conditions.ORB often failed to establish correspondence, while XFeat sometimes selected visually similar distractors.Source: B. Receiver-Side Target Acquisition… · p. 7
  • PATH projects the sender’s target location into the receiver image as a spatial prior, narrowing local acquisition when appearance is ambiguous.Appearance-based matching can still establish correspondence when the target remains visually distinctive.Source: B. Receiver-Side Target Acquisition… · pp. 7–8

C. Hardware Performance

PATH’s geometry-based handoff runs on CPU-only Jetson Nano hardware at video rate, using compact geometric messages rather than image or depth data. The reported feature-matching baselines use more compute and power and show less predictable timing on the same platform.

C. Hardware Performance

  • PATH’s CPU-only geometric operations run at 30–60 Hz on Jetson Nano without an additional GPU-based feature-matching stage.The GPU remains available for the onboard detector.Source: C. Hardware Performance · p. 8
  • Across 420 trials, bidirectional messaging averaged approximately 12 kB/s and sustained 30–60 Hz operation without communication-induced stalls.Packets were at most 128 B, with message latency well below the frame interval.Source: C. Hardware Performance · p. 8
  • ORB’s profile shows high CPU utilization, while XFeat adds frequent GPU bursts; both impose substantial compute and power loads on Jetson Nano.The profiles indicate limited headroom for other real-time tasks and timing variability on the embedded platform.Source: Figure 8 · p. 8, C. Hardware Performance · pp. 8–9, Figure 9 · p. 8

D. Mutual Agreement Handshake

PATH returns the receiver’s candidate to the sender for cross-view agreement before tracking responsibility transfers. The reported handoffs combine consistent commit timing after agreement with a two-UAV evaluation scope and stated implementation limitations.

D. Mutual Agreement Handshake

  • D2 observes D1’s fiducial, projects the received 3D target into its image, and returns a candidate for sender-side geometric agreement before transfer.Agreement must satisfy the acceptance criterion for the required consecutive frames before tracking responsibility transfers.Source: D. Mutual Agreement Handshake · p. 9, Figure 10 · p. 9
  • 7.20 s median and 7.95 s 90th-percentile time-to-handoff were measured across 42 successful handoffs out of 45 attempts.Three attempts timed out when sustained IoU could not be maintained within the allotted window.Source: D. Mutual Agreement Handshake · p. 9, Figure 11 · p. 9

VI. DISCUSSION AND FUTURE WORK

  • The implementation depends on reliable fiducial observation and communication, and the evaluation covers two-UAV handoff under tested indoor and outdoor conditions.The paper identifies markerless pose estimation, occlusion handling, and larger teams as future work.Source: §VI · p. 9

VII. CONCLUSION

  • Indoor geometric evaluation produced 0.047 m mean relative-position error and 0.030 m mean target-position error; ambiguous outdoor acquisition reached 96.0% accuracy.The conclusion reports 2.0% false-positive and false-negative rates and handoffs demonstrated on different UAV platforms.Source: §VII · pp. 9–10
Loading 2609.12456v1…