Research And Implementation Of UAV Detection System Based On Convolutional Neural Network

Consumer-grade uavs have low barriers to use and less control,and the phenomenon of heifei has become a problem that cannot be ignored.This paper studies the UAV detection system and analyzes the commonly used target detection algorithms and target tracking algorithms.The characteristics of KCF(Kernelized Correlation Filters)and the characteristics and response of context-aware tracking algorithm are improved.The UAV detection system is built.The main research work is as follows:1.Aiming at the target detection problem,we illustrate the traditional detection algorithm based on artificial feature and classifier,and the detection algorithm based on deep learning is further illustrated,the simulation verifies the recognition effect of the convolutional neural network used by the uav detection system.2.The problem of target tracking based on computer vision is studied.Aiming at the problem that the KCF tracking algorithm uses HOG(Histogram of Oriented Gradient)features,which makes the target response difficult to achieve pixel-level positioning.In this paper,the quadratic parabola model is used to approximate the discrete target response.The extremum of the continuous curve is obtained by Taylor formula.The extremum is used as the ultimate extremum of the target response.The tracking performance of the algorithm is verified on OTB50 dataset.Compared with KCF algorithm,the accuracy of this algorithm is improved by 0.13% and the success rate is increased by 1.4%.3.For the context-aware tracking algorithm,the traditional artificial features can not express the appearance of the target well.In this paper,the deep feature optimization algorithm is used to describe the appearance of the target.The confidence of the response is judged by the average peak correlation energy.It is verified on OTB100 dataset that the algorithm can improve tracking performance.Compared with STAPLE_CA(Sum of Template And Pixel-wise Learners_Context-Aware)algorithm,the accuracy is improved by 4.8% and the success rate is increased by 5.3%.4.Based on the research of target detection and target tracking,a uav detection system based on convolutional neural network is built.Darknet-based uav detection model is trained on Linux system.In this paper,the YOLOv2 target detection algorithm is transplanted on the VS2015 IDE,the hikvision PTZ(Pan/Tilt/Zoom)camera is used as video source,the system implements the conversion of the video format,and the system efficiency is improved by the double thread,the queue stores video,the mutex optimizes data access to critical resources,the GPU accelerates the detection,the PID(Proportion/Integration/Differentiation)controls the pan/tilt rotation.After testing,the system can realize the detection and tracking of the uav through the above operations,and achieve the expected effect.Compared with the unoptimized detection system,the detection speed of the system is increased by 78 times.