Hedef izleme algoritmalarıyla dağıtık füzyon için ilintisizleştirme yaklaşımları

The sensor data fusion can be described as merging complementary and/or overlapping information provided by sensors to produce superior information while ensuring consistency in every step of the process. The optimal fusion (centralized fusion) can be done when all measurements from the sensors are made available to fusion center. With modern real-world systems, due to communication constraints, the sensor data are usually first processed in local sensor/fusion nodes and the processed data is sent to other fusion nodes to be fused to produce better quality information. This approach is called track/estimate fusion. A fusion system where track/estimate fusion is made is called as a decentralised fusion system. The main problem of track fusion is the correlated estimation errors of the local agents. These errors originate from common process noise or common information due to previous communication which is not be easily identifiable. In addition, in noisy (i.e. false alarms, clutter) environment, decorrelated information can not be obtained precisely. Neglecting the correlations between the estimates to be fused has consequences ranging from low fusion performance to filter divergence due to overconfident estimates. In addition, the incompatible cases to be handled can occur during common information removal. This thesis proposes decorrelation algorithms to be employed in sensor fusion nodes in a decentralised fusion system using local IMM (Interacting Multiple Model), PDA (Probabilistic Data Association) or combination of these, IMM-PDA filters. Being instances of information decorrelation approaches, the proposed methods attack only the correlations caused by the previous communication. The basic idea is to recognize the common information and then remove it. The application of decorrelation has a significant effect on the filter performance in the fusion node. Information decorrelation is applied either on the Gaussian mixtures or the merged Gaussians produced by the IMM, PDA or IMM-PDA filters. The decorrelation approaches for the IMM, PDA and IMM-PDA filters are derived and analyzed in different chapters in the thesis along with the corresponding fusion operations. The implementation issues that might arise while applying the decorrelation approach are addressed in detail. The investigated approaches are compared with alternatives on simple distributed single (maneuvering) target tracking examples with various communication rates. Keywords: Information decorrelation, tracklet, distributed estimation, maneuvering target tracking, track fusion, interacting multiple model, IMM, probabilistic data association, PDA, IMM-PDA