The rise of FPV drones in modern warfare has created a need for effective active defense systems for infantry. Current solutions like electronic warfare and anti-drone guns face challenges due to frequency hopping and autonomous target acquisition. While manual shooting can be effective, it requires high skill and specialized equipment, which is not always available to soldiers in combat conditions.
The SMASH 3000 smart sight from the Israeli company SMARTSHOOTER is a known solution that assists in targeting. It allows a user to mark a drone, after which the system calculates the optimal firing moment based on distance, wind, and other ballistic factors. However, the system is primarily designed for stationary or slow-moving targets, prompting interest in developing more versatile alternatives.
A proof-of-concept project aims to create a compact, low-power device for real-time drone detection. The system focuses on identifying drones within a 30-meter range, ignoring background objects, and triggering a shot only when the target is centered. Key requirements include high detection speed, low power consumption, and the ability to operate for several days without heavy batteries.
The development process involved training a machine learning model on a dataset of 36 903 images, sourced from Kaggle and Roboflow. The data was split into 29 609 training and 7 294 validation images. To ensure accuracy, the developer used metrics like mAP50, mAP50:95, and center-point error, while also categorizing targets by distance to improve detection reliability.
Data cleaning was a critical step, involving the removal of duplicates and distorted images. The developer used perceptual hashing (pHash) to cluster similar images, ensuring that training and validation sets remained distinct. This approach helps in building a more robust model capable of distinguishing drones from other objects in various flight scenarios.
