Moving Objects Intelligent Extraction For Video Satellite On-orbit Real-time Service

With the rapid development of satellite remote sensing technology,the technology of earth observation from space has become an important manifestation of the country’s comprehensive strength.With the emergence of video satellite,the temporal resolution of earth observation data is improved to second level,which is more suitable for monitoring the dynamic changes of the target area.The application requirements have also shifted from the traditional periodic static census to the real-time monitoring of moving objects in key areas.Satellite videos have the characteristics of wide width and complex environment,the moving objects are small,lack of features,and low discrimination with the background.Compared with traditional surveillance videos,there is a new challenge to extract moving objects.The pseudo-motion caused by complex background interference with the detection of small-size moving objects.At the same time,the objects may be easily blocked in the complex environment,the tracking model update wrongly and the tracking fails,so the classical algorithm is difficult to be applied directly and effectively.The traditional satellite remote sensing application mode takes more time to respond to each link.In order to meet the needs of strong time-sensitive application,it is urgent to realize on-orbit real-time processing.Classical algorithms are difficult to process in real time under the condition of insufficient computing resources.Therefore,under the constraints of limited on-orbit computing resources,it is of great significance for the intelligent application of satellite video to detect typical moving objects on-orbit accurately,real time and intelligently,continuously track them and extract continuous dynamic information.The main contributions of this paper are as follows:1.Intelligent detection of moving objects in satellite video.Aiming at the problem of small objects in satellite video,much environmental background noise,and there are much pseudo-motion false detection in background,an intelligent detection algorithm based on adaptive background model and CNN is proposed,which combines the advantages of moving object detection and deep learning image object detection.Firstly,the background model is used to detect the motion foreground and generate motion candidate region.Then use an improved image classification CNN for motion candidate regions to detect,and accurately identify various objects and background noises.Finally,the classification results are fed back to the background model,and an adaptive local update mechanism is used to quickly update the false detection background to the model.The experimental results show that the F-Measure can reach 84%,and the accuracy is obviously improved.2.Intelligent tracking of moving objects in satellite video.Aiming at the problem that small objects in the satellite video is lack of features and complex background environment,and the object is easy to be completely blocked,which leads to the wrong update of the tracking model and the tracking failure,in order to enhance the robustness and accuracy of the tracking algorithm,an intelligent tracking algorithm based on correlation filtering and redetection is proposed.On the basis of the kernel correlation filter short-term tracker,an occlusion loss judgment mechanism is added to stop updating the model and activate the re-detection mechanism when tracking is abnormal.In order to accurately retrieve the object after it is reproduced,the high-dimensional feature expression of the moving object is realized and the space-time constraint is applied to achieve the long-term tracking.The experimental results show that the accuracy of the tracking algorithm is excellent,especially it can deal with complex scenes and show good performance.3.Video satellite on-orbit real-time intelligent detection and tracking moving objects system.For the problem that moving object detection algorithm usually has a large amount of computation and is difficult to be real-time under the constraints of computing capacity of spaceborne processing platform,use a combination of detection and tracking strategy to replace frame-by-frame detection with short-range tracking,and tracking tracklets are associated.On the premise of not reducing the accuracy of algorithm,the computation is reduced to improve the efficiency of processing on-orbit.Using satellite videos to conduct experiments on an embedded semi-physical simulation platform,the results show that the F-Measure of moving objects extraction can reach 83%,and the running speed reaches 20 fps,which can extract moving objects in real time with high accuracy and has good performance.