A Topology-Aware Spatiotemporal Handover Framework for Continuous Multi-UAV Tracking

When multiple drones track cars on a highway, they often lose track of a vehicle. This happens the moment it moves from one drone's view to another. This creates "trajectory fragmentation" (the loss of vehicle identity persistence). A single car is treated as a series of disconnected, anonymous objects.

Existing solutions usually try to solve this using Re-Identification (Re-ID). This involves extracting visual fingerprints to match a car in Camera A to a car in Camera B. But from a drone's top-down (nadir) view, most cars look identical. A white sedan is just a white blob. This paper argues that we should exploit the predictable physics of the road instead. By treating the multi-UAV setup as a logical pipeline, the authors achieve a Handover Success Rate (HOSR) of 99.8%.

The Problem

Current Multi-Camera Multi-Vehicle Tracking (MCMT) relies heavily on appearance-based Re-ID. This involves calculating high-dimensional visual embeddings (mathematical representations of appearance) to find matches via cosine similarity. This works for ground-level CCTV. There, you can see license plates or unique roof racks. It fails in aerial monitoring.

As shown in, nadir views suffer from extreme visual ambiguity.

Multiple vehicles of the same model appear indistinguishable. Illumination changes also pose a threat. Moving from sunlight into a building's shadow can radically alter a vehicle's perceived appearance.

Re-ID is also computationally expensive. Extracting these embeddings adds significant latency. This is a dealbreaker for real-time deployment on resource-constrained edge hardware (onboard computers with limited power). When drones act as isolated data silos, they produce disjointed "tracklets" (short, local trajectory segments). These fragments prevent higher-level network analysis, such as estimating origin-destination flows.

How It Works

The authors propose a shift from fragile visual matching to a deterministic, topology-aware handover mechanism. The architecture is visualized in .

It moves away from "looking" at the car. Instead, it predicts its path based on road geometry.

The system operates through three main stages:

1. Local Perception: Each UAV runs a high-throughput pipeline. It uses YOLO11 for detection and ByteTrack for local temporal association (linking detections over time). This keeps individual drone tracking stable during minor occlusions.
2. Directional Spatial Partitioning: The system discretizes the roadway into "Directional Zones" (e.g., Eastbound vs. Westbound). This acts as a coarse-grained filter. A vehicle exiting an upstream zone can only match a vehicle entering the corresponding downstream zone.
3. Deterministic Queue-Based Matching: This is the core contribution. The system uses a lateral-aware FIFO (First-In-First-Out) queue. When a vehicle enters a calibrated "Handover Zone," its metadata is pushed into a buffer. This zone is a non-convex polygon representing the overlap between two UAVs .

The metadata includes the ID, timestamp, and normalized lateral position ($y_{rel}$). When a new vehicle appears downstream, the system searches the buffer. It seeks a candidate that minimizes lateral displacement within a temporal window $\Delta t$.

The system focuses on where the car is relative to the lane and when it arrives. It bypasses the need to know exactly what the car looks like.

Numbers

The headline result is a massive leap in reliability. The authors report a Handover Success Rate (HOSR) of 99.8% in continuous traffic flows. This compares to just 74.1% for heavy-weight Re-ID baselines like FastReID. Even more impressive is the efficiency delta. While the Re-ID approach saw server-side throughput drop to 8.4 FPS (frames per second), the proposed method maintained 62.1 FPS on an NVIDIA V100 [Table I].

For engineers concerned with the "prod path" of edge deployment, the paper validates feasibility on the NVIDIA Jetson Orin NX. Using the YOLO11s model, the authors measured 25.6 FPS in 15W mode .

This is a critical finding. It suggests that a swarm of UAVs can each carry its own compute unit. They can maintain real-time, synchronized tracking without backhauling massive 4K streams to a central server.

What's Missing

The paper is technically sound, but three areas might cause production issues:

• Static Configuration Requirements: The handover zones in currently require manual annotation. In real-world deployments, UAVs experience wind-induced drift. Manual calibration is a scaling nightmare. The authors aim to automate this in future work.
• Congestion Sensitivity: The system struggles in high-density "stop-and-go" traffic (Set B). If a vehicle sits idle too long, it exceeds the static Time-to-Live (TTL) threshold. This causes the system to reset the identity. The system is optimized for flowing traffic, not gridlock.
• Nadir Assumption: The kinematic estimation relies on a linear Ground Sample Distance (GSD) coefficient. This assumes a strictly top-down view. If the UAVs are tilted, the mapping from pixels to metric meters becomes non-linear. This requires homography-based calibration (mapping planes between different perspectives).

Should You Prototype This

Yes, if you are building large-scale traffic monitoring systems. This applies to environments that are structured, like roads or tracks. The move from probabilistic visual matching to deterministic topological matching wins on both accuracy and latency.

This won't work for unstructured environments. For example, tracking pedestrians in a crowded park is difficult because there are no "lanes" to discretize. However, for any application where movement follows a predictable geometric path, the logic is highly reproducible. The source code is available at https://github.com/JYe9/multi-camera-multi-vehicle-tracking-system. Start by implementing the queue-based handover logic. The math is lightweight and should not blow your latency budget.

Figures from the paper