INVESTIGATION OF THE YOLOv5 ALGORITHM EFFICIENCY FOR DRONE RECOGNIZATION

Abstract

With the growth in the production and sale of drones, the number of offenses related to the use of drones in no-fly zones is increasing. Visual detection systems using neural networks for drone recognition effectively solve this problem. One of the most effective algorithms for objects of various classes is YOLO, which can be used to detect drones. However, this algorithm has a number of limitations that reduce the drone detection range. Therefore, the goal of the research is to evaluate the detecting drones' effectiveness at different distances using the fifth version of this algorithm, as well as to create algorithms for increasing the detection range. Based on the experimental data, datasets were created for training four main modifications: s, m, l, and x of the neural network of the YOLOv5 algorithm. These network modifications were trained for the visible and infrared (IR) ranges, as well as various image resolutions at the network input. After processing a dataset of drone flight videos, the effectiveness of various modifications of the neural network was evaluated, the dependences of the probability of detecting a drone on the distance and speed of the drone were plotted, and the maximum detection range was estimated. A two-stage algorithm has been created that makes it possible to increase the detection probability and increase the detection range as a result of a combination of the classical YOLOv5 algorithm at the first stage of processing and the convolutional neural network proposed by the authors at the second stage. An algorithm for synthesizing IR images has been created to supplement IR datasets with the necessary drone-background combinations when training neural networks. Practical recommendations are given for choosing the type of neural network and quantitative estimates of the YOLOv5 algorithm's efficiency in combination with a two-stage processing algorithm.