Research On Track-before-detect For Weak Target And Track Fusion

With the opening of low-altitude airspace and the diversification of small-sized aircrafts,weak targets have become the new growth point of economic development,but increasingly sudden intrusions show that weak targets have also been significantly potential threats for civil aviation,key areas,important activities,and even Homeland security.The trend to prevent and control weak targets has become increasingly severe.Limited by the low signalto-noise ratio(SNR)characteristics of the target echoes,both traditionally thresholding detection for short-term coherent accumulation and single-frame tracking filter cannot meet the requirements for weak targets.Additionally,prolonged accumulation time may bring limited signal-to-noise ratio gain,but it in turn would cause that the target moves crossresolution units within signal frame.By contrast,multi-frame joint processing can not only estimate the continuous motion state,but also improve the detection SNR by accumulating gain among frames.So it becomes an effective technology for weak target prevention and control.The track-before-detect(TBD)technology,as a classical multi-frame joint processing method,has attracted worldwide attention in recent years for its remarkable achievements in weak target detection and tracking.However,it is found that there are still some problems in engineering application,such as inadequate measurement error,long detection delay and poor tracking accuracy,which need to be solved urgently.At the same time,as the distributed radar network can expand detection range,improve detection performance and system stability,it is also an effectively technical means to improve the ability of weak target prevention and control.However,the problem of fusion state estimation accuracy under the framework of interval communication is also a thorny problem to be solved.Therefore,this paper carries out related research work around the above issues.The main contents and research results are summarized as follows:1.For traditional DP-TBD algorithms,errors in amplitude and mapping dimensions existed simultaneously in engineering are not fully considered.Therefore,a dynamic programming TBD(DP-TBD)algorithm based on joint log-likelihood ratio test(JLLRT)is proposed.The proposed method draws lessons from the probability data association(PDA)which takes prediction measurement as the center and weights the candidate mapping dimension measurements in correlation gate according to the distance differences.Without changing the definition of admissible transitions,confidence areas centered at located cells are established.The mapping dimension measurements of adjacent cells are modeled as detection probabilities conditioned on that the target moves at located cells,and the amplitudes are modeled as the likelihoods under the assumption that it is the echo of the target.Weighting the likelihoods of cells in confidence area by detection probabilities results in signal-frame merit.The sum of merits along an admissible transition is the joint loglikelihood ratio of the transition.All transitions are traversed and thresholded according to the Neyman-Pearson criterion.Once the target is detected,its continuous states are filtered by inputting mapping measurements of the resultant transition.In this way,errors in the amplitude and mapping dimensions existed simultaneously are considered comprehensively.Different simulations indicate that the proposed method is superior to the generalized likelihood ratio test DP-TBDs in detection probability,tracking probability,estimation accuracy and running speed.2.For traditional PF-TBD algorithms,weighting all particles to estimate target state gives rise to poor tracking accuracy and long detection delay before particles aggregate.Therefore,a particle filter TBD(PF-TBD)algorithm based on particle inheritance and backtracking is proposed.Fortunately,according to the continuity of particle inheritance and backtracking among different frames in importance resampling,the closer the particle is to the target,the more sub-particles it would generate.On the contrary,the particle far away from the target would not generate sub-particles continuously.Therefore,we propose to retrospect parent particles at different moments of each sub-particle at detection moments,and estimate the historical states only by weighting the parent particles retrospectively.In this way,it is effective to avoid the influence of disperse particles on the accuracy of state estimation before all particles aggregate.Additionally,more precise estimations promote to detect target quickly.The simulations explore the gains of such a method in detection probability and estimated accuracy comparing with traditional PF-TBDs.3.As the moment that the target appears is unknown,the detection window with fixed length would not declare the presence of the target until the gain among frames meet the requirement of detected SNR in sliding batch processing.Therefore,the design of telescopic detection window with fixed end is proposed.To detect the target timely,the proposed design fixes the detection end,and extends the window length frame by frame.Once detection statistics for windows of different length exceed any grading threshold,the target would be detected promptly.Combining such a design with the proposed PF-TBD would further shorten the detection delay.The simulation results show that the design can effectively improve the detection performance.4.To solve the problem that existing decorrelated track fusion algorithms are limited to full-rate communication architecture or measurement errors that need to be known globally,a decorrelated track fusion algorithm for distributed network is proposed.The proposed algorithm follows the Kalman filter architecture,and globalize process noise in local state prediction.Therefore,local filter states are decorrelated,and weighted according to the inverse of information matrixes at fusion moment.From the Bayesian point of view,the fusion state is globally optimal.Both the simulations and the measured data show that the tracking performance of the proposed algorithm is better than that of the tracklet fusion algorithm.