Enhanced Least Squares Support Vector Machines for Decision Modeling in a Multi-Sensor Fusion Framework

In this article we introduce a software framework for embedded online data fusion on different levels of data abstraction. We present our data oriented fusion model and introduce the main functional units. The paper is focused to the decision modeling process. In our approach we use Support Vector Machines (SVM) as well as Least Squares SVM (LS-SVM) for decision modeling. Due to the computation complexity and the necessary memory requirements we prefer LS-SVM for the classification tasks. The main disadvantage of LS-SVM is the loss of sparseness by using equality constraints instead of inequality constraints in the cost function. We introduce a novel method for intelligent data preselection (PTD LS-SVM) to compensate for this short coming. Experimental results demonstrate the feasibility of this approach.