Moving Vehicle Object Detection And Tracking Based On Satellite Video

At present,the traditional ground spatial information acquisition methods have been difficult to meet the demand for intelligent sensing and real-time tracking of moving targets in the fields of intelligent transportation,public security,resource monitoring and disaster monitoring.With the development of satellite remote sensing technology,satellite video data with minute-level revisit period,second-level time resolution and sub-meter level spatial resolution provide data support for intelligent sensing and real-time tracking of moving objects.Because of the difference between the satellite video and the ground video,the moving target detection and tracking algorithm for the ground video cannot be well applied to satellite video.Therefore,the research of moving target detection and tracking algorithm applicable to satellite video is an important problem that needs to be solved urgently.In this thesis,the vehicle,a typical moving object,was selected as the research object to study the rapid detection and intelligent tracking method for satellite video vehicle.However,the current research still has the following shortcomings: It is found that background subtraction method was widely used in the detection of moving vehicles in satellite video,but there was still a problem that background subtraction method had not been evaluated comprehensively in satellite video,which led to the restriction of in-depth research of background subtraction method in satellite video;The moving satellite platform,illumination change and other factors led to the pseudo-motion moving targets in satellite video,and the background subtraction method was not good at distinguishing the false moving target,which led to the decrease in the accuracy of vehicle detection;The intelligence of target tracking in satellite video was not high and the existing intelligent tracking methods had low robustness and were prone to model drift when subject to the common challenges of target tracking,such as occlusion and background interference.In view of the above problems,the main research contents of this thesis were as follows:(1)Aiming at the problem that the background subtraction method had not been comprehensively evaluated in satellite video,the detection performance of 42 background subtraction methods for moving vehicles in satellite video was evaluated,and the most suitable background subtraction method for satellite video was summarized.Firstly,42 background subtractions provided by BGS Library were used to detect moving vehicles on three satellite videos.Then,in order to improve the detection results,a general detection framework was used,including image pre-processing,background subtraction parameter optimization and postprocessing of the detection results.Finally,the detection results were evaluated and analyzed using accuracy evaluation metrics.(2)Aiming at the problem that the background subtraction method had poor ability to distinguish pseudo moving targets,a moving vehicle detection method,MOGv1 x,which could effectively distinguish pseudo moving targets in satellite video was constructed.Firstly,trajectory accumulation on the foreground mask extracted by the basic background subtraction method ABL was proposed,and the pseudo moving objects in the foreground mask after trajectory accumulation were removed to obtain the region of interest of the vehicle.Then,under the constraint of the region of interest,the background subtraction MOGv1 based on multiple Gaussian models was used to detect moving vehicles using the general detection framework.(3)Aiming at the problems of low intelligence of target tracking methods in satellite video and low robustness of existing intelligent tracking methods,a vehicle intelligent tracking method,Siamese RPNx,which could effectively deal with the challenges of occlusion and background interference in satellite video was constructed.Firstly,Siamese RPN was introduced for intelligent tracking of vehicles.Then,the robustness of Siamese RPN was improved using four strategies: bounding box size update,inertial mechanism-based bounding box penalty,motion foreground enhancement,and Occlusion Detection and Processing.It is shown that,(1)MOGv1 has the highest detection accuracy in the satellite video vehicle detection dataset,The generic detection framework used in this thesis can effectively improve the detection accuracy of different background subtraction methods.(2)The moving vehicle detection method MOGv1 x proposed in this thesis is basically free from the interference of pseudo-moving targets,and its accuracy rate is greatly improved compared with MOGv1.(3)The vehicle tracking method Siamese RPNx proposed in this thesis can effectively deal with the challenges of occlusion and background interference,and achieve higher tracking accuracy compared with other common trackers.The method evaluation carried out in this thesis provides ideas for the improvement of background subtraction.The proposed detection and tracking method of moving vehicle has achieved good results on satellite video.The above work could provide a reference for improving the accuracy of moving target detection and the robustness of target tracking based on satellite video.