DISTRIBUTED ESTIMATION IN DISTRIBUTED SENSOR NETWORKS (MULTI-TARGET TRACKING, INFORMATION FUSION)
Date of Completion
Engineering, Electronics and Electrical
The assessment of dynamic situations using data from multiple sensors occurs in many military and civilian applications. In some applications, the data are collected by a network of sensors distributed over a large geographic region. In such distributed sensor networks (DSN), because of considerations such as reliability, survivability and communication bandwidth, centralized processing is either undesirable or infeasible. Instead, the sensors supply data to a set of local processors/nodes which are connected by a communication network. The nodes process the local sensor data and exchange processing results with other nodes. This distributed processing architecture creates the problem of how to combine (fuse) the local processed results from multiple sensors into a global one. The scope of this research is to study and develop an estimation algorithm for such a system.^ In practical estimation problems, difficulties come from the uncertainties of measurement origins and system models. In this research, these two important problems are studied in detail in a distributed framework. The concept of redundant information is introduced. A general fusion algorithm which combines the local processed results into a global one is developed. The basic distributed estimation technique has been applied together with the Probabilistic Data Association (PDA) scheme to handle the data association problem when measurement origins are uncertain. The multiple model concept has also been employed for state estimation of system with uncertain models.^ Fusion algorithms are derived for (1) the Joint PDA, (2) Interactive Multiple Model (IMM) and (3) the PDA with Multiple Models algorithms. We show that the necessary condition for the fusion algorithms to be optimal is that the frequency of communication is equal to the frequency of measurement. We also describe the sufficient information to be communicated for each algorithm. If the frequency of communication is less than the frequency of measurement, the latest local estimates (or conditional pdfs) are no longer sufficient to construct the global estimates. This is primarily due to the model switches and the propagation of the common process noises between communication times. A suboptimal procedure for the latter case has also been obtained. ^
CHANG, KUO-CHU, "DISTRIBUTED ESTIMATION IN DISTRIBUTED SENSOR NETWORKS (MULTI-TARGET TRACKING, INFORMATION FUSION)" (1986). Doctoral Dissertations. AAI8629915.