Research And Application Of Moving Target Detection Algorithm Based On Compressed Sensing Domain Video

Moving target detection is the key technology in the bottom layer of intelligent video surveillance system,which lays a foundation for the video understanding and intelligent analysis of the upper layer.It has a wide range of applications in the field of video surveillance involving public safety.Classical moving target detection methods mainly deal with spatial domain or compressed domain video.These methods are effective and meet the real-time requirements,but there are still some shortcomings.One is that it is difficult to adapt to the surveillance scenes in complex situations,the other is that the cost of capturing,transmitting,processing and storing video data is high.More robust and less expensive data processing methods have yet to be proposed.Therefore,based on the compressive sensing theory,this paper studies the detection of moving target based on the Compressed-Sensed-Domain video.The main work of this paper includes:(1)A moving target detection method based on background subtraction and tensor decomposition is proposed to solve the moving target detection problem of Compressed-Sensed-Domain video under different scenes.This paper studies the more a priori information of the video background in the space-time structure under the tensor framework,that is,the "differential low-rank property" of the video background in different modes and different scenes.In this paper,we first construct the initial video decomposition model based on the background difference,and then use the tensor Tucker decomposition to model the video background and introduce "adaptive rank constraint update" for Tucker decomposition.Finally,an improved background model is obtained.The new method consists of two parts:video reconstruction and detection of moving targets.The experimental results show that this method not only effectively improves the accuracy of moving target detection in Compressed-Sensed-Domain video under different scenes,but also further improves the reconstruction accuracy of video frames.(2)A video reconstruction method based on M-estimation is proposed to solve the video reconstruction problem when the video signal is mixed with impulsive noise in the process of compressing and sampling.In this paper,Welsch M-estimation is introduced as a cost function to measure the video reconstruction error.Combined with the previous background model,a compressed sensing domain video reconstruction and moving target detection method can be obtained,which can effectively reduce the impact of impulsive noise.The experimental results show that the new method has good robust performance in reconstruction performance when the video signal is mixed with noise with different signal-to-noise ratio or impact strength during compressive sampling.(3)A software simulation system of moving target detection based on Compressed-Sensed-Domain video was designed and implemented.The video reconstruction and moving target detection simulation under various monitoring scenes and different noises were carried out.The system can fully demonstrate the video signal compressive sampling,reconstruction and moving target detection and performance evaluation of all the processes.After the whole process is finished,we can see the results of reconstruction and moving target detection of Compressed-Sensed-Domain video,as well as the evaluation index for these results:peak signal-to-noise ratio,structural similarity and comprehensive evaluation index.