Research On UAV Target Detection And Formation Recognition Based On Computer Vision

With the rapid development of intelligent processing,self-organizing communication and other technologies,various countries are vigorously developing unmanned clusters,which are characterized by autonomous collaboration,large scale,low cost and unmanned,and are more and more widely used in the military fie ld.Due to its small size,low altitude flight and other characteristics,it is not easy to be detected by radar,which often leads to false detection and missing detection,which brings great challenges to our air defense.Therefore,it is one of the main research directions to reduce the airspace threat brought by UAVs to conduct target detection,confirmation and early warning,identify the formation of UAVs and then take corresponding countermeasures such as interception.Along with the rise of computer vision technology,it has advantage of higher recognition accuracy to detect the target for UAVs in detection of video frames and identify formation formation by using deep learning to deal with image frames which is captured by the target detection device in the airspace.However,there are some problems in UAVs target detection and formation recognition using conventional deep learning based on computer vision: 1)The neural network algorithm occupies a large amount of memory space;2)Due to the influence of UAVs target characteristics and light intensity,the accuracy of detection and recognition cannot meet the requirements;3)There is no large amount of open formation database for training.According to the above problem,in this paper,by comparing the mainstream convolution neural network algorithm,Single Shot-box Detector algorithm which has advantages in small target detection was improved,and construction of neural network algorithm is used to identify the formation.The following studies are mainly carried out:The detection algorithm model which uses MobileNet to replace the Visual Geometry Group trunk network in the Single Shot Multi-box Detector algorithm is transplanted to mobile devices to realize real-time detection of airspace targets.Activation functions in the MobileNet network are easy to cause neuronal degeneration.In order to greatly reduce the amount of convolution calculation on the mobile terminal,reduce the memory occupation,improve the detection speed,and ensure the timeliness of UAVs warning,MobileNetv2 is adopted as the backbone network by improving the activation function.The original Single Shot Multi-box Detector uses shallow feature map to detect small targets,which is easy to miss detection of small targets such as UAVs.Therefore,in this paper,the deep feature maps which have rich semantics and the shallow feature maps which have rich details are fused into the pyramid network structure.At the same time,the size of inputed image is enlarged to improve the semantic information of the shallow feature map,and the loss function is improved to achieve the convergence state of the model,so as to reduce the error of the predicted value of the model to improve the classification accuracy.Homomorphic filter is used to preprocess the influence of light intensity in data concentration.The improved algorithm in this paper can effectively improve the detection accuracy of small targets such as UAVs.Matlab simulation was used to build the formation of the UAVs.The formation control algorithm based on Flocking obstacle avoidance was used to avoid collision between individuals,maintain the consistent speed and the stability of topological structure,which could quickly adjust the movement state of the formation and achieve the simulation test of the formation control flight of the UAVs group to establish the formation database of the UAVs.The classification and recognition algorithm of convolutional neural network is constructed for training and recognition to improve the recognition accuracy of the formation of the UAVs.