- Case study ↓
- AI/ML
- Computer Vision
Aircraft Detection That Spots Targets Just a Few Pixels Wide, in Time to Avoid Collisions
A deep learning system that finds distant aircraft and drones, as small as 3–5 pixels, in high-resolution video filmed from a moving aircraft.
The demo
See Drone Detection at work
- Industry
- Computer Vision
- Input
- High-resolution aerial video
- Output
- Annotated video, FastAPI
- Built with
- YOLOv5 · Swin Transformer · PyTorch · FastAPI
The problem
Distant aircraft are almost invisible on camera
Seen from another aircraft, a plane or drone a long way off can be just 3–5 pixels wide. The camera is moving too, so the sky, clouds and ground shift in every frame and ordinary detectors either miss the target or raise false alarms.
For aviation safety, a late warning is no warning. Spotting other aircraft early is what makes collision avoidance possible.
The solution
What Drone Detection does
Tiny-target detectionFinds aircraft only 3–5 pixels wide
Works from a moving cameraHandles shifting sky, clouds and ground
Spatio-temporal modelUses motion across frames, not single images, to cut false alarms
Annotated video outputEvery detected aircraft boxed through the whole clip
Retraining pipelineThe model can be retrained as new footage comes in
Technology
How it’s built
Every step, from the high-resolution aerial video to the results.
Step 1
Input
- Aerial video
- Label Studio
Step 2
Detection
- YOLOv5
- Detectron2
- OpenCV
Step 3
Temporal analysis
- Swin Transformer
- PyTorch
Step 4
Output
- Annotated video
- FastAPI
- Detection accuracy on aerial video samples
- 85%+
- Inference time per frame
- 0.3 s
- Size of the smallest targets detected
- 3–5 px
Who it’s for
- Aviation and UAV makersAdding vision-based detect-and-avoid to their aircraft.
- Defence and research teamsWho need small-target detection in aerial footage.
- Airspace security providersCombining detection with their existing camera systems.
Case study
Drone Detection: the full case study
- Overview
An AI solution that detects extremely small moving objects, such as distant aircraft just 3–5 pixels wide, in wide-angle, high-resolution video filmed from a moving aircraft. The goal is to improve aviation safety by giving early warning of other aircraft and reducing the risk of mid-air collisions.
- Problem
Detecting objects only 3–5 pixels wide is hard when the camera itself is moving and the background changes in every frame. Existing methods were unreliable, so the client needed an automated AI system for early warning.
- Solution
ML Bench built a detection system that combines deep learning with temporal analysis. It takes video in, runs the model on every frame and returns the video with each detected aircraft boxed.
The system includes a custom training and retraining pipeline, with footage annotated in Label Studio, and is served through FastAPI.
- Challenges & solutions
YOLOv8 looks at one frame at a time, so it struggled with tiny moving objects.Moved to a spatio-temporal approach that understands motion across frames.
The client needed reliable detection on a moving, changing background.Used a Swin Transformer that processes video sequences, improving detection of small moving objects in aerial footage.
- Results
- 85%+ detection accuracy on aerial video samples
- Inference time optimised to 0.3 s per frame
Have Questions?
Find the right solution
for now
We tried. Frame-by-frame YOLO models struggle with targets this small on a moving background, so we added a Swin Transformer that reads motion across frames.
Why ML Bench
Why teams choose ML Bench
Built for real-world footageWe build computer vision for real cameras and scans: retail, sports, healthcare, logistics and safety.
Starts with the data you haveExisting cameras, recordings or image archives are usually enough for a first proof of concept.
Custom to your setupWe adapt the model to your camera angles, image quality, classes and workflow, not the other way round.
One team, pilot to productionData, model training, dashboard, API and deployment handled by the same team.
Next AI/ML project →
Theft Detection
AI video analytics that reviews retail CCTV footage, scores suspicious customer behaviour and timestamps likely theft.
Free assessment
See What AI Spots in Your Footage
Send us a short aerial clip and we'll show you what the model detects, free.
- UI/UX Design
- Web Development
- Mobile Apps
- SEO
- AI/ML Solutions
