Вакансия Computer Vision Engineer (Miltech)

Skills

• OpenCV
• Jetson
• Raspberry Pi
• ARM SBC
• Python
• C++

Описание вакансии

We’re developing the world’s smallest and most affordable missile designed specifically to counter Russian FPV drones – a hand-launched, camera-guided system that can be carried in multiples by Ukrainian soldiers as a critical last line of defense. With successful test flights already underway and a basic onboard vision system in place, we’re now looking for a skilled Computer Vision Engineer to help take this technology to the next level. Your work will focus on advancing detection, filtering, and tracking capabilities in complex, real-world conditions using embedded, real-time computer vision. This is more than just a technical challenge – it’s a chance to make a direct impact on Ukraine’s defense. We’re also in the process of establishing a local presence in Ukraine, with the longer-term goal of forming a joint Norwegian–Ukrainian defense company that combines global expertise with strong local talent and leadership.

Requirements:

• Strong experience with classical computer vision techniques (OpenCV or similar).
• Solid background in real-time vision pipelines for video streams.
• Experience working with gyroscope or IMU data to stabilize motion.
• Familiarity with small neural networks for classification or filtering.
• Ability to optimize performance on embedded systems (Jetson, Raspberry Pi, ARM SBCs).
• Experience integrating multiple CV stages into a single real-time loop.
• Deep understanding of false positive reduction, robustness, and temporal filtering.
• Proficiency in Python or C++ (ideally both).
• Upper-Intermediate English level.

Responsibilities:

• Use gyroscope data to estimate camera motion and stabilize frames.
• Remove the horizon and background using a lightweight neural network.
• Detect drone candidates using fast visual features in the foreground.
• Filter detections with a small neural network that runs on image patches.
• Merge nearby detections into single targets.
• Track valid detections across frames using visual and geometric consistency.
• Trigger re-detection if tracking fails, reusing the stabilized view.
• Run periodic confirmation using a stronger classifier on locked targets.
• All of this must run in real time on embedded hardware with very limited compute – and still be robust to motion blur, low light, trees, birds, and background clutter.