Design Of UAV Intrusion Detection System Based On Machine Vision

The rapid development and wide application of civilian drones has not only brought convenience to life,but also led to incidents that violated personal privacy and threatened public security.UAV intrusion detection based on visible light technology has become a research hotspot in the field of machine vision due to its high dependence on rapidly developing machine learning technology.However,there are still some problems to be solved in UAV intrusion detection system based on visible light technology,which includes the small pixel proportion of in the image and its tendency for UAVs to melt into a complex context.Based on the demand of UAV intrusion detection system and the principle of keeping abreast of research hotspots,this paper designs a UAV intrusion detection system based on machine vision.The research includes the following:(1)In view of the fact that the Vi Be(Visual Background extractor),a kind of moving target detection algorithm,is prone to Ghost when detecting targets whose moving state changes,an improved Vi Be algorithm based on the median method background compensation is proposed.The median method background compensation is to use the median method to establish the background model before the Vi Be algorithm background modeling,avoiding the possibility of Ghost when the moving target is used as the background element in the first frame of the image.In addition,the median method background modeling is performed once every 500 frames of images,using 50 frames of images.When the target motion state changes,the historical multi-contrast mechanism is used to determine the UAV moving target,which avoids the possibility of Ghost when the UAV target is gradually absorbed into the Vi Be background.And a dispersion coefficient is added for adaptive foreground and background segmentation to adapt to light intensity changes in different scenes.Through multi-algorithm comparison,the improved Vi Be algorithm’s accuracy rate has increased by 4%,and the recall rate has increased by 8%.(2)31,683 UAV target recognition samples are obtained through real-time shooting of three UAVs at different circumstances including angles,scales,and attitudes,as well as network acquisition and sample expansion.And he training set and test set are constructed at the ratio of 9:1.The target recognition algorithms including YOLOV4(You Only Look Once version 4),YOLOV3 and SSD(Single Shot Multi Box Detector),which all belong to one-stage,are trained and tested on the same sample set.The average accuracy of YOLOV4 was 5.18% higher than that of YOLOV3 and 4.8% higher than that of SSD.The average recall rate of YOLOV4 is 3.21% higher than YOLOV3 and 2.17% higher than SSD,so YOLOV4 is selected as the target recognition algorithm in this paper.(3)A Kalman filter-based STC(Spatio-Temporal Context)target tracking algorithm is proposed against the phenomenon of increasing error Spatio-Temporal Context information under the similar background with UAV.It extracts the spatial context information from the ROI(Single Shot Multi Box Detector)that is determined to contain the moving target of the UAV under the sky background,as well as the temporal context information of the target in the continuous video sequence.However,the Kalman filter is used to locate the target in a complex background,which avoids the possibility of superimposing the Spatio-Temporal Context information through STC algorithm.Through comparative analysis,the improved algorithm’s average tracking success rate has increased by 10.3% compared to the original algorithm,and the average center position error has been reduced by 6.2 pixels.(4)Combined with the mentioned above target detection algorithm and target tracking algorithm,this paper sets up UAVs intrusion detection system with the client through the PTZ(Pan/ Title/ Zoom),and carries out target detection characterized by the PTZ tilt and zoom lens and target tracking featured with PTZ tilt.All that can make us fulfill real-time detection of UAVs invasion.