IAMARS: drone detection, tracking and trajectory prediction

YOLOv8n finds small drones in video, ByteTrack keeps a stable ID for each one, a Kalman filter estimates position and velocity, and a constant-velocity model predicts the next half second. Every number on this page was produced by a script in the repository.

Code on GitHub Model card Model weights

Four drones against open sky

Pipeline output: each drone keeps its ID (1 to 4) for the whole clip. Boxes show track ID and confidence, lines show the recent path, orange marks the predicted path. Input clip supplied by the project author.

Pipeline

Video frame→ YOLOv8n detection→ ByteTrack IDs→ Kalman filter→ Trajectory prediction

Detection

Test set	Images	mAP@0.5	mAP@0.5:0.95	Precision	Recall	Source
VisioDECT held-out test (block split)	1,800	0.965	0.596	0.960	0.957	json
Anti-UAV-RGBT frames, never seen in training	3,424	0.291	0.143	0.571	0.274	json

The second row measures the domain gap: the model was trained only on VisioDECT.

Speed

Device	Mean latency / frame	95th percentile	Pipeline FPS	Source
NVIDIA GeForce RTX 2050	11.3 ms	15.16 ms	88.5	json
13th Gen Intel(R) Core(TM) i5-13420H (CPU only)	49.51 ms	56.85 ms	20.2	json

Detect + track + Kalman + predict per frame, batch 1, 1280 × 720 video, laptop on AC power.

Tracking

Evaluation	Videos	MOTA	IDF1	ID switches	Source
Anti-UAV-RGBT test, IAMARS detector + ByteTrack	91	0.067	0.178	452	json
Same videos, ground-truth boxes (tracker upper bound)	91	0.928	0.485	1,384	json
MultiUAV swarm videos, ground-truth boxes	20	0.985	0.924	297	json

MOTA and IDF1 from py-motmetrics (match at IoU ≥ 0.5). With perfect boxes the tracker keeps identities well on static-camera swarm videos but not on Anti-UAV, whose pan-tilt camera makes boxes jump.

Second example: 24 drones in formation

Pipeline output on a swarm. Source clip: "MagicLab – 24 Drone Flight" by Marco Tempest, CC BY 3.0, via Wikimedia Commons, trimmed and annotated.

Intercept-point prediction (simulation)

Simulation only. 1,000 trials per pattern; 3D position measured at 30 Hz with 1.0 m noise; after 2.0 s a Kalman filter estimate is used to aim a 80 m/s interceptor. Hit = within 1.0 m of the true position.

Flight pattern	Hit rate (≤ 1 m)	Median miss	Hit rate with true state
Straight	28.3 %	1.36 m	100.0 %
Turning	1.4 %	13.52 m	8.2 %
Jinking (random manoeuvres)	0.0 %	16.65 m	0.3 %

Source: results/intercept_simulation.json

Run it yourself

git clone https://github.com/aviraj1805/intelligent-aerial-monitoring.git
cd intelligent-aerial-monitoring
pip install -r requirements.txt
python scripts/download_assets.py
python app.py            # interactive web demo at http://127.0.0.1:7860

This page is static because hosted Python apps on Hugging Face require a paid plan. The interactive Gradio app runs locally with the commands above.