| In recent years,China has faced an intricate national security situation,the task of air defense has bare the heavy burdens.The air defense early warning and countermeasures have become more and more difficult.Therefore,a series of new ground-to-air weapon systems have begun to be installed in troops.The complete use of traditional live-ammunition training methods cannot meet the requirements,not only the targeted training and improvement are difficult to achieve,but also the actual needs of the effectiveness of military expenditure cannot be maximized.The research on supporting training systems has a great influence on targeted training and saving military expenditures.During the training process,the triggering of training operation instructions and the evaluation of training results are based on the detection of aircraft targets.The low-altitude aircraft targets are in the sky image which are captured by the mobile camera can be tracked,detected and identified by video detection technology.The study can be divided into the following parts:(1)The detection of Moving target.The background motion of the video screen was caused by camera motion.It is difficult to extract moving targets accurately by traditional target detection methods.This paper proposes a method based on region scoring and motion compensation for moving target detection,and put forward fragment integration based on HSI(Hue、Saturation、Intensity)color space and motion consistency Method.Compared with other algorithms,the detection accuracy of this algorithm has improved significantly in cloudy and half-air backgrounds.(2)The identification of target type.In order to identify the type of moving target accurately on the basis of moving target detection,this study analyzes the R-CNN(RegionConvolutional Neural Networks)series algorithm based on the candidate area and the YOLO(You Only Look Onec)series algorithm based on regression.The development of the YOLO V3 for the needs of small target recognition mainly manifested in target detection of a larger-scale feature maps and the import of residual network in the detection network.From the Establishment of aircraft target data set and the improvement of network training,it can be concluded that the network’s ability to recognize small targets has improved significantly through the comparison between the improved YOLO V3,unimproved YOLO V3,Faster RCNN algorithm experimentally.(3)The tracking of target.The single-frame processing speed of the moving target detection and target recognition algorithm cannot meet the real requirements,thus a tracking algorithm can be used to replace the detection and recognition work partly.In this paper,the comparison experiment of DSST(Discriminatiive Scale Space Tracker),KCF(Kernel Correlation Filter),TLD(Tracking Learning Detection),Camshift(Continuously Adaptive Mean-Shift)tracking results show that the DSST algorithm single target Tracking frame rate can reach more than 95 FPS,and this method also can adapt to the target’s scale changes.(4)The Integration of system.This paper introduces the system integration ideas and framework.In order to improve the overall processing speed,a tracker can be used to track the target of the subsequent sequence based on the detection and recognition results of a few frames.The process included the acquisition of tracking targets,the creation and update of tracker,the deletion of departure target tracker,the discrimination of flight trend,etc.The experimental results show that the detection accuracy of the integrated system is more than81%,and the processing frame rate is between 18 ～ 35 FPS,which basically meets real requirements. |
PDF Full Text Request