Research On The Technology And Application Of Moving Target Detection Based On Compressive Sensing

Moving target sensing is one of the hottest direction in the field of information acquisition in recent years.Compressive sensing combines sampling and compression into one,eliminating the operation of collecting a large amount of redundant data and then compressing it in the high-speed sampling process.It can alleviate the pressure of high-speed sampling.At the same time,it effectively reduces the impact of packet loss and error coding during data transmission.This makes it possible to obtain high spatial resolution information by low-cost detectors.This dissertation deeply studies the key technologies of target detection and tracking based on compressive sensing and its applications.The detailed research content is as follows:For the case of a large-scale target changing into a point target in a complex background,an adaptive algorithm-switching strategy is proposed,which combines the Fast Compressive Tracking(FCT)algorithm with Kalman filter.And the way to determine the threshold value in the process of algorithm switching is described in detail.At the same time,two target detection algorithms are selected according to the target size to achieve multi-scale detection of small targets.Experimental results show that this method can achieve real-time tracking when target size gradually changes.In order to improve the robustness of FCT to light change,target rotation,motion blur,etc.,an improved algorithm is proposed.The algorithm extracts target features from Gaussian difference maps and works only in the region of interest(ROI)to reduce the complexity of algorithm.A naive Bayesian classifier is used to obtain N candidate targets with N highest classifying scores.Then weighted cosine similarities between these candidate targets and the initial ground truth target and the most similar target in current frame are calculated to determine the target in the current frame.Experimental results show that the proposed algorithm outperforms other tracking algorithms(such as CT,FCT,CSK,CXT,CPF,Struck and CNT).And its average frame rate can reach 50.8 frames per second,which can meet the real-time requirement.To omit the image reconstruction and improve the efficiency of small target detection based on compressive sensing,a small moving target detection algorithm based on template matching in compressed domain is proposed.This algorithm obtains the foreground measurements by background modeling in compressed domain.The simulation results show that when the target moving distance within a single sampling interval is less than 0.18 units,target motion almost has no effect on the final positioning accuracy.Meanwhile,the influence of downsampling factor,projection matrix,number of measurements,noise,etc.on the detection performance is analyzed.As the increase of downsampling factor will lead to the drop of detection rate and false-alarm rate simultaneously,it is proposed to reduce the downsampling factor of region of interest to improve the target detection rate,and increase the global downsampling factor to further reduce the amount of data while controlling the false alarm rate.Experimental results show that when the number of measurements is 2(the global downsampling factor is 36,the ROI downsampling factor is 4,and the size of the ROI is 18 x 18),the detection rate can reach 0.960;the false-alarm rate is only 0.009;the compression ratio ranges from(12.26:1)to(15.46:1);and the single-frame detection time is only 0.34 seconds.A good trade-off between detection performance and runtime can be achieved.In order to reduce the sampling time and calculation amount of reconstruction,a parallel compressive imaging system based a digital micromirror device is studied in depth.Based on this,a parallel compressive imaging system based on a set of masks are designed.After that,the proposed small moving target detection algorithms in compressed domain are improved according to the characteristics of the measurements collected by the desktop mockup.Experimental results show that when the global downsampling factor is 36,the ROI downsampling factor is 4 and the size of the ROI is 18x18,the detection rate can approach to 1,and the false-alarm rate can be kept at a low level.This method can effectively reduce the system cost,and can obtain high-resolution target location information directly from low-resolution measurements.However,this system cannot be applied to the detection of high-speed small targets whose moving speed is faster than 0.773 units/s.