Anti-Drone Detection & Interception for PRG
Real-time computer vision to detect, track, and intercept hostile loitering-munition drones before they reach their target
Evaluated, Vetted, and Delievered v2.2 for use by the client effective Mar. 2026
Case Study: Anti-Drone Detection & Interception for PRG
Background
Low-cost loitering munitions like the Shahed and Geran have upended the economics of modern air defense. A single drone costing tens of thousands of dollars routinely forces the launch of interceptors costing tens of millions—an unsustainable exchange rate that exhausts defensive infrastructure far faster than it can be replenished. Closing that gap starts with a cheaper, faster, and more reliable way to see the threat coming.
Challenge
Spotting a small, fast, low-flying drone against a cluttered sky in potentially inclimate weather conditions is a genuinely hard computer-vision problem. Targets are tiny and distant, frequently appear in fog, darkness, or thermal imagery, and are easily confused with birds and aircraft. A practical system has to run in real time, miss almost nothing—a single missed drone is a mission failure—and rarely cry wolf, all while staying cheap enough to deploy at the scale the threat demands.
Solution
PRG Dynamics and the Peregrine system take both names from the fastest animal on earth—a bird of prey built to hunt other flyers. We trained a custom computer-vision model to do the same: swiftly detect and track down hostile drones the instant they appear, before handing them off to an interception mechanism. Three things seat it comfortably above the rest:
Exhaustive 1.8M Image Dataset
From a single set of 4,651 images to a unified corpus of ~1.8M images across 60+ SOTA datasets, our curated dataset spans Shaheds, Gerans, and Bayraktar variants, anti-UAV RGB-thermal pairs, fog, dusk and night-vision, distant sub-pixel targets, and live field tests from Pakistan to Turkey—plus rejection imagery of birds, aircraft, people, vehicles, and power lines.
State-of-the-Art (SOTA) Model
Iteratively developed and refined from an initial YOLOv8m pre-trained on industry-standard legacy data to a tailor-made YOLO26m arcitechture: 20.4M parameters, 78.7 GFLOPs, and a ~52 MB footprint, running at 13.7 ms per frame: the PRG model not only builds upon the industry's best—it sets a new standard in accuracy and efficiency.
Hyper-Optimized Training
We trained our model for months on end using Stochastic Gradient Descent (SGD), a cosine learning-rate schedule, warmup, and mosaic augmentation before the final set of stablizer epochs. A custom fitness function then overrode YOLO's default to prioritize catching every drone in view while boxing it tightly enough to safely intercept before reaching its target.
Under the Hood
- Threat + rejection classes: The hostile "drone" class is the target; helicopter, airplane, bird, person, and six vehicle/infrastructure classes exist purely so the model can reject them rather than fire on anything that moves.
- Recall-first fitness: YOLO's default checkpoint score does not prioritize recall enough for our use case; Peregrine's re-weighting makes a missed drone the most-penalized outcome, then rewards tight boxes to minimize collateral damage (if any).
- All-conditions coverage: visible, low-chroma/thermal, fog, dusk, and night-vision imagery, with distant small-target and real field-test footage to ensure detection in any real-world environment.
- Frame-rate-independent timing: Optical Character Recognition (OCR) on burned-in video timestamps ensures the drone stays in view, even if the camera feed is choppy.
- Low-weight & Accessible: a ~52 MB model at 13.7 ms per frame, suitable for efficient and immediate deployment to any operating system at a desired point of defense.
Outcome: Real-time detection that holds up in the real world.
These are drone-class detection numbers, and they must be read against the data behind them. Version 1.0 scored near-perfect on a single, narrow Shahed set it could all but memorize—but graded retroactively on the common held-out test set every later version faces, it collapses to 17.4% mAP@50, exposing how little it actually generalized. Every version from v1.3 on is measured on that same held-out set—approximately 387,000 images never seen during any training run—drawn from a corpus of ~1.8M images across 60+ SOTA datasets spanning fog, night-vision, thermal, distant targets, live field tests, and twelve rejection classes. Beginning with v2.0 (February 2026), the backbone was upgraded from YOLOv8m to YOLO11m. Scoring was re-weighted to prioritize recall and tight localization over the leaderboard metric YOLO optimizes by default. A v3.0 built on YOLO26 is currently undergoing optimization with initial results below.
Drone-Class Detection Performance by Version
| Version | Dataset | Precision | Recall | F1 | mAP@50 | mAP@50-95 |
|---|---|---|---|---|---|---|
| v1.0 | 4,651 | 1.2%† | 55.9%† | 2.3%† | 17.4%† | 17.0%† |
| v1.3 | 355,500 | 89.0% | 48.3% | 62.6% | 65.7% | 35.0% |
| v2.0 | ~518K | 92.5% | 85.2% | 88.7% | 89.0% | 52.2% |
| v2.3 | ~1.8M | 94.8% | 88.2% | 91.4% | 92.0% | 53.6% |
Drone-class metrics. All six versions were evaluated using the same held-out v2.3 test set (~387,000 images, never seen during any training run). Dataset column = total training-corpus images across train + val + test splits.
Peregrine v2.3 detecting a Shahed in real time on unseen test footage.
The clip runs the v2.3 model on live Shahed footage it never trained on, at real-time speed. Detection is proven; the next phase closes the loop—turning each detection into a steering signal that guides an interceptor onto the target.
Contact Us Today
Have questions regarding our software or services? Reach out to our team and let's discuss how we can help bring your ideas to life.
