Multi-Target Detection And Tracking Techniques For Distributed Through-the-Wall Imaging Radar

Through-the-wall imaging radar（TWIR）emits low-frequency electromagnetic waves to penetrate through walls,and obtain real-time locations of human targets in the building,which has important application values in urban combat,anti-terrorism and stability maintenance,disaster rescue and other fields.Due to the complexity of the indoor environment,single radar detection has encountered many difficulties such as blind area,low accuracy,and false target interference.Distributed through-the-wall imaging radar（DTWIR）deploys multiple TWIR nodes in multiple viewpoints to cooperatively detect indoor targets,which could effectively overcome the above shortcomings.In recent years,DTWIR has become a research hotspot.This dissertation addresses the problem of robust indoor human target tracking by DTWIR.The research works mainly focus on target detection,target tracking and information fusion for DTWIR.The main contributions are listed as follows:1.A multi-target likelihood ratio detection algorithm based on the Gaussian mixture model is proposed.Firstly,the amplitude distributions of targets and background noise are modeled as Gaussian mixture distribution.Subsequently,the parameters of the Gaussian components are estimated based on the expectation-maximization algorithm.The proposed algorithm could adaptively detect multiple extended targets under unknown background noise.2.An extended target detection algorithm based on fully convolutional network and attention mechanism is proposed.The network consists of a ”downsampling-upsampling”structure and jump connections integrated with attention gates,which could utilize both amplitude and shape features to detect multi-scale extended targets.Moreover,the proposed algorithm has the ability to distinguish multiple adjacent targets.3.A shape-adaptive extended target tracking algorithm is proposed based on the Kalman filter and the mean-shift algorithm.Firstly,the target state is predicted based on the motion model.Subsequently,the shape-adaptive mean-shift algorithm is utilized to find the target region.Finally,the Kalman filter is utilized to estimate the target state.The proposed algorithm could address the tracking problem of multiple shape variant targets.4.A global track estimation algorithm based on Hankel matrix completion is proposed for the track broken problem caused by object occlusion.Firstly,the track association hypotheses are established based on time and velocity constraints.Then the complete track in each hypothesis is estimated using the Hankel matrix completion technique.Finally,the optimal association hypotheses are selected via generalized linear assignment.The proposed algorithm could simultaneously associate the short tracks and estimate the missing track.5.A radar registration algorithm based on the Gaussian mixture model is proposed,which calculates the relative positions of the radar nodes based on the detection or tracking results of each node.Subsequently,two fusion algorithms are proposed for robust multitarget tracking in complex indoor scenarios,including the sequential multi-sensor joint probabilistic data association based measurements fusion algorithm and the covariance intersection based track fusion algorithm.