Hawk Eye

Hawkeye

Inspiration

It all started when we began noticing news headlines that most people scroll past without a second thought. A drone is killing thousands of innocent children in war in Iran. A drone shut down Gatwick Airport for three full days. A drone spotted near Heathrow's runway, forcing flight diversions and passenger chaos. Smugglers are using consumer drones to carry contraband across borders under the cover of night. Also in this era, drones are also used for attacking with explosives. Things we have seen in recent wars.!!!

As we are getting more and more advanced, protecting our privacy is becoming a threat, anyone can keep an eye on us and passing confidential information without even knowing. Existing detection systems were failing — not because drones are invisible, but because they look almost identical to birds from a camera due to the same size, same altitude, and similar speed. Surveillance systems would flag a flock of pigeons as a threat, operators would get flooded with false alerts, and eventually they would stop trusting the system entirely. The drone industry grew to $24.72 billion in 2020 alone, putting sophisticated flying cameras into the hands of everyday people at a scale nobody had prepared for. So, we wanted to build something that did not just detect objects in the sky but also told the right person immediately.

What it does

Hawk Eye is a real-time drone detection and alert system. I will must say we people get tired but our ai never get it is 24/7 protecting and giving information. You point a camera at the sky, and Hawk Eye watches it for you. The moment it sees something airborne, it classifies it as either a drone or bird or anything suspicious looking like missiles or thus, if it is a drone, it immediately sends an SMS alert to the designated operator. The operator does not need to be watching a screen. Hawk Eye watches it for them and calls them when it sees drones, actually. *The inference pipeline can handle three types of input: a static image, a pre-recorded video, or a live camera stream. * Each frame runs through the model, which returns a bounding box and a confidence score. If the confidence crosses our threshold and the class is drone, the system triggers the SMS alert through Twilio. We designed the alert pipeline to run asynchronously, so it does not slow down the inference loop — detection and alerting happen in parallel, not in sequence. This makes Hawk Eye useful across a range of real environments. At airports, it filters genuine drone threats from the constant background noise of bird activity, so shutdowns happen only when they need to. At border checkpoints, it can autonomously monitor crossing points through the night when drone-based smuggling is most active. In residential and corporate settings, it can catch surveillance drones before they capture a single frame of private footage. And on autonomous delivery drones themselves, onboard detection means a drone can identify a bird in its flight path and reroute automatically.

How we built it

We built Hawk Eye around YOLOv4 — You Only Look Once, version 4 — as our core detection model. We used Google Colab for training the model; the dataset was prepared by ourselves. We used Roboflow for creating and managing the dataset. For sending data we used Twilio..

Challenges we ran into

The hardest part was collecting the dataset and labeling it properly. Also, we need to wake so late at night to train our model, and we have to make configuration edits to customize the training and get best mean average precision.. We labeled almost 3000 images, which was a lot of work to do. It was our first time building an ml project, we had no idea where to start from, what to do, how to make it work properly. We trained the model three times to improve accuracy, it has really tiring.

Accomplishments that we're proud of

It was our first time doing an ml project, we had no idea how to start off, what to do. But at the end we are proud to build a model that actually works. Truly, we didn't expect that we will be able to do this far. We are proud of what we did.

What we learned

We learned that data is the hardest part of computer vision. Diversified and lot of data is actually needed for building a good working model. We also learned to manage a team and work together. To train the model to differentiate between a bird and a drone was literally the hardest part, our model can do it with a good amount of accuracy.

What's next for Hawk Eye

The most immediate step is upgrading the model from YOLOv4 to a newer architecture like YOLOv8 or RT-DETR, to push accuracy and inference speed further, especially for small objects at altitude where the classification is hardest. We also want to package Hawk Eye for edge deployment on hardware like Raspberry Pi or Jetson Nano, so it can run as a low-cost, low-power installation at border posts, remote facilities, or residential properties without needing a full server setup. And we want to extend the alert system beyond SMS- push notifications, email, API webhooks so Hawk Eye can slot into existing security infrastructure rather than sitting beside it.