Contemporary drones are becoming more and more intelligent thanks to the integration of deep learning and computer vision. In the last study published in IPSJ transactions in computer vision and applicationsInnovative methods of detecting and tracking drones using multi -American systems and advanced algorithms were presented.
Scientists proposed a system consisting of a static wide -angle camera and a rotating tower with a narrow -angle camera in high resolution. This system can detect small objects at long distances and analyze them in detail with zoom cameras. At the basis of this technology, the Yolov3 architecture is modified optimized for quick and accurate detection.
The system presented in the study consists of several key elements, each of which plays a key role in ensuring high accuracy and efficiency of drone detection and tracking.
Wide -angle camera
The wide -angle camera is mounted on a stationary platform with a 16 mm focal length lens, providing a field of view of about 110 °. This allows the camera to cover large areas and monitoring over long distances. The camera can send images in a resolution of 2000 x 1700 pixels at a speed of about 25 frames per second. A wide field of view plays a key role in the initial detection of small objects, such as drones on the horizon, enabling the system to quickly respond to new objects in its field of view.
A rotating tower with a narrow angle
The second camera in the system is mounted on a rotating tower, which allows it to change the field of view and track objects detected by a wide -angle camera. A narrow angle camera is equipped with a 300 mm focal lens, providing a field of view of about 8.2 ° and the ability to enlarge over 35 times. This camera has been designed for a detailed analysis of objects over long distances, enabling the system to thoroughly identify and track drones. The tower can quickly rotate and adjust the camera angle to capture high -quality images of target objects.
Main computing unit based on Linux
The central element of the system is the main computing unit, which is a Linux -based computer equipped with a NVIDIA graphic processor. This unit processes images captured by cameras and performs deep learning algorithms on a graphics processor that is Nvidia GeForce K620 with 2 GB of memory. The use of GPU allows the system to process large data volume in real time and ensures high efficiency in performing complex computational tasks. Deep learning algorithms such as Yolov3 have been modified and optimized to work in this system, achieving high accuracy and detection speed.
These components operate in strict integration to ensure high reliability and performance of drone detection and tracking system. The interaction between wide -angle and narrow cameras, along with a powerful computing unit, allows the system to quickly and thoroughly respond to the appearance of drones in its field of view, ensuring a high level of safety and control.
To detect drones, the system uses a modified version of Yolov3, which efficiently processes images and detects small objects. Unlike standard approaches, this method uses the regression model for the quick location of objects in the images. The modification of Yolov3 architecture included a reduction in the number of filters while maintaining the number of layers, optimizing the system operation on limited GPU resources.
This innovative system ensures high accuracy and detection speed, which makes it ideal for use in safety and supervision tasks. He is able to detect drones on the horizon, follow their movements and, if necessary, analyzing them using a narrow angle camera.
Studies show the potential of deep learning in increasing the capabilities of drones to perform complex tasks in real time, including supervision, safety and rescue operations.
