Topology-Aware Multi-Camera Tracking (JYe9)
Table of Contents

⚠ Disclaimer: This entry may be incomplete, out of date, or inaccurate. It is AI-maintained on a best-effort basis. Do not rely on it as a sole source — verify claims independently using the sources listed below.

Summary

This project solves the inverse of this section’s usual problem: it tracks ground vehicles across multiple UAV-mounted camera feeds, rather than tracking drones from fixed ground cameras. It’s included here as an adjacent, adaptable reference architecture — the core technique (assigning one consistent identity to a moving target as it crosses between camera fields of view, using geometric/topological reasoning rather than pure visual re-identification) is the same cross-camera correlation problem that a fixed multi-camera drone-detection array has to solve. Each of three UAV video sources runs YOLO detection and ByteTrack per-camera tracking; a “topology-aware spatiotemporal handover” framework then assigns global IDs to vehicles as they move between camera overlap regions, described in an associated paper submitted to ICUAS 2026.

Key Facts

  • Author: JYe9 (GitHub)
  • Type: Open-source software + associated paper — cross-camera vehicle tracking (UAV-sourced video, not drone detection itself)
  • License: Academic and Non-Commercial Research License — not permissively licensed; commercial/production reuse would require separate arrangement
  • Status: 33 stars, 0 forks as of this review
  • Detection/tracking: Ultralytics YOLO (custom 3UAVs.pt model) + ByteTrack via the Supervision library, per camera
  • Cross-camera matching: Global ID correlation within defined camera-overlap regions
  • Associated paper: “A Topology-Aware Spatiotemporal Handover Framework for Continuous Multi-UAV Tracking” — per the repository’s own citation, submitted to ICUAS 2026 (arXiv:2605.15779 per the repo; this entry could not independently confirm the arXiv listing’s content — the abstract page did not return readable text at time of review)

How It Works

The system processes video from three UAV camera sources in parallel. Each stream runs an Ultralytics YOLO model fine-tuned for vehicle detection, feeding ByteTrack (via the Supervision library) for within-camera tracking. As a tracked vehicle moves from one camera’s field of view into an overlapping neighbor’s, the topology-aware handover framework — using known geometric relationships between camera coverage areas rather than relying solely on visual appearance matching — reassigns the same global ID rather than spawning a new track. Output includes per-camera annotated video, plus CSV metadata with per-vehicle speed (via pixel-to-meter calibration) and direction, frame index, global ID, camera ID, and status.

Why This Is Here (Not a Drone-Detection Tool)

This is UAV-as-sensor, not UAV-as-target — the opposite configuration from the fixed-camera-array C-UAS use case this section otherwise documents. It’s included because the specific sub-problem it solves — persistent identity across overlapping camera fields of view without continuous appearance-based re-identification — is directly relevant to any multi-camera drone-detection deployment, including TRAMIS’s own cross-camera track correlation (which uses timestamp synchronization plus heading-closeness rather than a topology/handover model) and the general multi-sensor fusion problem this section prioritizes. Treat it as a technique/architecture reference, not a candidate C-UAS product.

Related / Further Reading (Not Independently Verified)

The following were flagged as adjacent references but could not be independently confirmed for this entry — treat as pointers for further research, not verified facts:

  • OpenVINO Multi-Camera Multi-Target Tracking demo (openvinotoolkit/open_model_zoo) — reported to combine detection with re-identification embeddings for cross-camera association (person/vehicle-oriented, described as a reusable architecture pattern). Direct fetch of this repository was blocked (robots.txt) for this entry; current status, license, and maintenance activity are unconfirmed.
  • Kinematics-based multi-camera aerial-target re-identification research (e.g., work on “Multiple Aerial Targets Re-Identification by 2D- and 3D-Kinematics-Based Matching”) — described as graph-matching approaches to cross-camera re-ID requiring little or no appearance-based training data, conceptually adjacent to the trajectory/kinematics-based (not appearance-based) approach used by TRAMIS. Specific paper(s) not identified/verified for this entry.

Limitations

  • Non-commercial license: The “Academic and Non-Commercial Research License” is a real constraint — this cannot be dropped into a commercial C-UAS product without separate licensing, unlike the MIT/permissively-licensed projects elsewhere in this section.
  • Not drone detection: Targets are ground vehicles; sensors are airborne. Applying the technique to the drone-detection use case (fixed ground cameras, airborne targets) would require nontrivial adaptation, not a drop-in swap.
  • No published cross-camera accuracy metrics (ID-switch rate, handover accuracy) in the repository as of this review — the associated paper (unconfirmed access) may contain these.

Sources

← prev next →