Research On UAV Target Detection Algorithm Based On Video

With the rapid development of Unmanned Aerial Vehicle(UAV)technology,also known as drones,the application of drone technology has been widely used in our daily lives.However,while promoting the rapid progress of human society,it also brings hidden dangers and challenges to our social security and even military security.The “black flying” and “indiscriminate flying” of drones seriously damage public safety and personal privacy,so it is very important for the research of anti-drone systems.Drone detection technology is a crucial first step in the anti-drone system,which provides the necessary preconditions for the subsequent measures to eliminate hidden dangers.The drone is a typical low slow small target which has the characteristics of small sizes,variable flight conditions,complex flight environments.In order to effectively detect small UAV,this paper chooses a method based on video detection to analyze the detection problems of drones.This paper mainly detects and recognizes the target of the four-rotor small UAV in the video image,and designs two different detection methods for the camera which has two states: fixed camera and moving camera.Since there are no open drone data sets,in order to effectively validate and train the designed model,we have built our own drone data sets,which include various challenges: different weather conditions,complex background changes,shaking branches,changes in lighting,occlusion of low-altitude buildings,interference from other moving objects(especially birds),drone motion blur,and multiple motion states at different distances from the camera.For the detection of drones under fixed cameras,we obtain the region of interest through the moving target detection algorithm,and then extract the Fourier descriptor(FD)and the Histogram of Oriented Gradient(HOG)feature for the region of interest.The final FD+HOG feature is sent to the trained Support Vector Machine(SVM)classifier for classification and identification.In addition,in order to verify that the FD+HOG+SVM algorithm can more accurately identify the drone,we conducted experiments on the accuracy of classification and identification of drones and non-drones,drones and birds.The results show that the algorithm can still effectively identify drone images on small data sets.The combination of moving target detection algorithm and traditional machine learning can effectively detect the UAV,this greatly reduces the computational complexity.While taking into account the calculation speed,the recognition rate of the drone can be further improved to meet the real-time demand.For the detection of drones under moving cameras,We introduce the candidate region-based two-stage detection algorithm represented by R-CNN series and the end-to-end single-stage detection algorithm represented by YOLO.Based on the YOLOv3 model,we have made some improvements from the size of the input image,the size of the anchor boxes,and the selection and fusion of the multi-scale features of the network.Finally,the experimental comparison and analysis of Faster R-CNN,YOLOv3 and improved YOLOv3 are carried out,which proves that the improved network can detect the small target of drone more accurately and quickly.