In this article
TL;DR:
- Drones with onboard AI now process computer vision locally, without cloud dependency
- Applications range from precision agriculture to search and rescue with thermal detection
- NVIDIA Jetson Orin Nano and dedicated chips are democratising edge AI access
- The EU AI Act imposes clear limits on facial recognition in drones
- Professional solutions starting at around €2,500
Two years ago, talking about artificial intelligence in drones meant, in most cases, simple object-following algorithms or pre-programmed flight modes. Today we’re in a completely different reality. Drones with embedded AI processors are making real-time decisions, identifying objects, people and patterns without sending a single byte to the cloud. And this changes everything.
In this article, I’ll explore the current state of computer vision in drones, the technologies behind it, the practical applications already transforming entire sectors across Europe, and the ethical and regulatory challenges this advancement brings.
What exactly is onboard AI in a drone?
When we say “onboard AI” (or edge AI), we’re referring to a drone’s ability to process artificial intelligence algorithms directly on the hardware it carries, without needing internet connectivity or remote servers. It’s the opposite of the cloud-dependent model that dominated the first generation of “smart” drones.
In practice, the drone carries a neural processor — typically an NPU (Neural Processing Unit) chip or mobile GPU — capable of running machine learning models in real time. And when I say real time, I mean inference at 30-60 frames per second, with latency under 50 milliseconds.
Why does this matter? Imagine a search and rescue drone flying over a forested area after a wildfire. If it needs to send every video frame to a remote server, process it, and receive a response, we’re talking about latencies of 200-500ms with good 5G coverage — and much more in rural areas without it. With local processing, detecting an injured person happens instantaneously.
The technologies behind it: chips and frameworks
The chip ecosystem for drone AI has evolved dramatically. Here are the main players in 2026:
NVIDIA Jetson Orin Nano
The Jetson Orin Nano has become virtually the industry standard for professional AI drones. With up to 67 TOPS (Tera Operations Per Second) since the Super update of December 2024 (40 TOPS in the original version) and power consumption between 7 and 25 W, it offers an impressive performance-to-power ratio. The developer kit costs 249 US dollars, which makes it accessible even for academic projects.
In benchmarks published by Ultralytics, the YOLO26n model optimised with TensorRT (FP16) runs on the Orin Nano Super in 4.57 milliseconds per image, more than sufficient for real-time object detection. The Portuguese company Beyond Vision, for example, uses an NVIDIA Jetson module as the brain of its HEIFU hexacopter.
Qualcomm Flight RB5
Particularly interesting for smaller consumer drones. It integrates AI processing, 5G connectivity (with an optional modem) and flight management on a single board. The DJI Mini 4 Pro and other consumer drones also already do advanced tracking onboard.
Hailo-8
The Hailo-8 is particularly impressive — 26 TOPS at just 2.5W power consumption, in an M.2 form factor that fits almost any platform. It can serve, for example, in thermal camera drones for real-time thermal and visual data fusion.
Practical applications that are already reality
Let’s move beyond theory and look at what’s actually happening.
Precision agriculture
In agriculture, drones with multispectral cameras and onboard processing can identify water stress, pests and nutritional deficiencies during the flight itself.
For crops such as vineyards, olive groves or citrus, the AI model needs to be trained with data from those crops.
Infrastructure inspection
In Portugal, Labelec, part of the EDP group, inspects power lines with helicopters and drones and has developed a machine learning model that automatically classifies the data from those inspections. REN is also testing drones, and stork nests are a real problem in Portugal: over 20 years, storks were the second biggest cause of disturbances on REN’s lines.
Enterprise drones like the DJI Mavic 3 Enterprise offer some of these capabilities out of the box, but the most advanced solutions use customised platforms with multiple cameras and Jetson processing.
Search and rescue
This is perhaps the most exciting application — where onboard AI genuinely saves lives. In Portugal, some civil protection services already use drones that detect victims by the heat they give off, as the Matosinhos municipality announced in April 2026.
In a search, a drone with a thermal camera and onboard detection quickly covers areas that would take ground teams hours to search.
Computer vision models running on drones
For the more technically inclined, here’s what’s actually running on these drones:
- YOLO (You Only Look Once): The YOLO family dominates object detection, from YOLOv8 to YOLO26, the latest version (January 2026). Nano and small models run easily on embedded hardware
- SAM (Segment Anything Model): Edge-optimised versions (MobileSAM, FastSAM) enable real-time semantic segmentation
- Depth estimation: MiDaS and variants allow drones to “perceive” distances without LIDAR
- Onboard OCR: For reading plates and serial numbers during inspections
The clear trend is quantisation — reducing weight precision from FP32 to INT8 or INT4, maintaining 95-98% of original accuracy while gaining 3-5x speed. Frameworks like TensorRT, ONNX Runtime, and Hailo Dataflow Compiler make this increasingly accessible.
European regulation and ethical questions
The EU AI Act, in force since August 2024 with its prohibitions applying since 2 February 2025, bans real-time remote biometric identification in publicly accessible spaces for law enforcement purposes, save for limited exceptions. In other words, real-time facial recognition from police drones is, as a rule, prohibited in the EU, and other uses remain subject to the GDPR. And rightly so, in my opinion.
This doesn’t prevent generic person detection (without identification), object detection, or aggregated behavioural pattern analysis. But it draws a clear line that the sector needed.
For professional AI drone use in the EU, these are the obligations that matter most:
- Operator registration with the national aviation authority, mandatory for drones weighing 250 g or more and for drones with a camera that are not toys
- GDPR compliance whenever the images allow people to be identified
- A data protection impact assessment (GDPR Article 35) when publicly accessible areas are systematically monitored on a large scale
Costs and accessibility in 2026
Rough cost ranges:
- DIY/academic solution: Drone + Jetson Orin Nano + camera — from €1,200
- Professional inspection kit: €2,500-5,000
- Enterprise solution: €8,000-25,000 (multi-sensor, certified)
- Consumer with basic AI: from €239 (DJI Neo 2, Mini 4 Pro, etc.)
For those wanting to start experimenting, I recommend a 4K drone for beginners as a base platform, potentially adding a Hailo or Jetson module for AI projects.
Conclusion
Onboard computer vision in drones isn’t the future — it’s the present. It’s transforming real sectors, creating jobs, and raising regulatory questions we need to keep debating.
For those in the industry, the message is clear: investing in AI and computer vision skills isn’t optional. It’s the baseline for remaining competitive over the next five years. For curious outsiders, there’s never been a better time to get in — costs are low, hardware is accessible, and demand for qualified professionals is enormous.
Related Articles
- DJI Mavic 3 Enterprise: is it worth the investment?
- Thermal camera drones: worth the investment?
- Investing in thermal cameras for drones
- Hybrid drone technologies
Frequently Asked Questions
Do I need programming skills to use AI drones?
It depends. Consumer drones like the DJI Neo 2 or Mini 4 Pro come with pre-configured AI that works out of the box. For custom professional solutions, Python knowledge and frameworks like TensorRT are recommended.
Does onboard AI work without internet?
Yes, that’s precisely the advantage. Processing occurs locally on the drone with no internet required. Results can be stored locally and synced later. The exception is AI model updates, which require prior download.
Is it legal to use AI drones for surveillance in Europe?
Generic person and object detection is legal for authorised professional purposes. Real-time facial recognition in publicly accessible spaces for law enforcement purposes is prohibited under the EU AI Act, save for limited exceptions. Flying requires operator registration with the aviation authority and, outside the Open Category, an operational authorisation; when the images identify people, the GDPR applies.
What’s the most affordable AI drone to start with?
For pre-configured AI, the DJI Neo 2 (€239 for the drone only, on DJI’s European store) offers capable AI tracking. For custom projects, a kit with Jetson Orin Nano starts at €1,200 — but requires significant technical knowledge.
DroneFlyBlog
Keep exploring the world of drones
Get new buying, aerial photography and drone use guides to prepare for your next purchase or flight.