Labeled Random Finite Set Based Distributed Fusion Over Sensor Network

To deal with the threat of weak objects and the challenge of complicate surveillance environment,a new type of surveillance systems,namely,the sensor network is proposed.In practice,both the observations and the objects in the surveillance scenario are of highly uncertain in the sense that the characteristics of the points and their number are both unknown and time-varying,as a result,it is difficult to describe and model them in the framework of conventional statistical theory.As a promising advanced signal processing technology,the random finite set based multi-source multi-object fusion technology is one of the front line gambits in the international signal processing and information fusion communities.By modeling all the uncertainties including randomness,inaccuracy,fuzziness of observations and an unknown and time-varying number of objects in the framework of point process theory,this technology provides a unified description and modeling approach for the multi-source multi-target fusion problem.However,as a newly developed technology,it is still faced with a lot of problems and challenges,for instance,the problem of non-standard observation model,the unknown level of correlation among different sensors,the distributed fusion of labeld random distributions,etc.To solve the aforementioned problems,this thesis investigates the labeld random finite set based multi-object filtering problem under the generalized observation mdoel(GOM),the label inconsistency problem among sensors,the robust and efficient distributed fusion algorithms with labeled random distributions.The main contributions of the thesis are as follow:1.In many applications,there might be sensor observations,such as superpositional sensors,merged measurements,and object occlusions,etc.,which cannot be accurately modelled using the standard multi-object likelihood.To this end,this thesis proposes a multi-object filtering algorithm under GOM using labeled random finite sets,called the labeled multi-object(LMO)filter for GOM(LMO-GOM).With sufficient computing resources,the LMO-GOM filter is expected to exhibit the optimal performance in principle,and possibly served as a performance benchmark in labeled multi-object filtering problem.Further,fast implementation methods adopting principled approximations are proposed,which can reduce the computing burdern significantly with a slight performance loss.2.This thesis formulates the “label inconsistency” among different sensor from the perspective of information divergence.Further,the impact of “label inconsistency” among different sensors on the performance of generalized Covariance Intersection(GCI)fusion is analysed in a principled way.The analysis shows that the GCI fusion with labeled densities is highly sensitive to inconsistencies between label information embedded in different local sensor nodes.3.To counteract the “label inconsistency” sensitivity problem,this thesis proposes a label-free GCI(LF-GCI)fusion for labeled random finite sets.Through the projection of the GCI fusion on the target kinematic state space,the LF-GCI fusion can achieve robustness to the “label inconsistency” phenomenon,while providing an enhanced fusion performance.4.Considering the computationally inefficiency of LF-GCI fusion as well as the “label inconsistency” sensitivity problem,this thesis proposes a label matching GCI fusion(LM-GCI)for labeled random finite sets.By formulating the optimization problem of label matching as a linear assignment problem,the problem can be resolved using some fast solvers.Hence,the proposed method is robust to the label inconsisteny problem as well as computionally efficient.