Research On Visual Recognition Technology In Anti-UAV System

With the rapid development of UAV,people's lifestyle has changed a lot.However,the abuse of UAV become more and more common,which has brought great security risks to the whole society.Therefore,how to effectively supervise UAVs has become a hot issue around the world.To this end,we propose an Anti-UAV system based on visual detection and recognition.Firstly,a wide-angle camera is used to detect objects in the sky that are suspected of UAVs.Next,the telephoto camera is used to identify and track the detected targets.At last,the radio suppression is used to control the UAV until the danger alarm is eliminated.In this thesis,the key technologies in the Anti-UAV system are divided into four parts: small moving target detection,image super-resolution,small object recognition and object tracking.The main research contents are as follows:For the detection of small moving targets,we propose an adaptive threshold Gaussian mixture model.Compared with the traditional threshold-fixed Gaussian mixture model,the biggest advantage of this method is that it can adaptively adjust the threshold according to the complexity of the background.Therefore,the target can be fully detected in a simple background,and the noise in the complex background is suppressed as much as possible.This method has stronger anti-interference ability against complex background and can effectively improve the robustness of the target detection algorithm.For small object recognition,we combine the super-resolution network with the traditional object recognition network.We propose a novel recurrent convolutional network,which considers the fixed-point iterative algorithm as the forward propagation of the recurrent network and replaces the nonlinear layer in the recurrent network with a convolutional neural network.Therefore,this method can be regarded as a master of classical numerical analysis theory and deep learning.In addition,this method can achieve different degrees of super-resolution effects by controlling the number of recurrent layers,and has strong theoretical value and application value.Experiments show that this method can alleviate the low accuracy problem to some extent.For the object tracking,we introduces the Actor-Critic framework in the reinforcement learning into the object tracking,and optimizes the long-term tracking as a whole problem.These methods can bring convenience to the UAV tracking algorithm.The experiences of the four video in real shot show that the accuracy of the proposed method is better than that of the Kernel Correlation Filter.In addition to testing the results on the dataset,we also did a lot of experiments and deployed the entire system in a nuclear power plant.The whole system is still in good condition.