Using Support Vector Machine to Monitor Behavior of an Object Based WSN System

Abstract

In this paper, we propose a new method to classify the bound error values of Kullback-Leibler Distance (KLD)-particle filter (PF) based Support Vector Machine (SVM) to reduce the mean number of particles used (sampling) as well as improve the performance of runtime in reality for monitoring an object. In wireless sensor network (WSN) system, the object location is calculated via the collected received signal strength (RSS) variations which are effected by furniture, walls or reflections. Therefore, we propose an architecture diagram to track an object and build the dataset model. By transforming the system state model from the 1D to 2D, the bound error value of KLD resampling can enhance estimation accuracy and convergence rate of declining number of particle used by generating a sample set near the high-likelihood region for ameliorating the effect of the RSS variations. Our proposal considers how to classify and find the bound error values of KLD PF for each iteration. The first iteration, using the observation information via KLD resampling optimal bound error to conduct a resampling on the basis of the initial bound error. From the second to the end iteration, we propose the SVM technique to search the predicted bound error value that fulfills the minimum of mean number of particle used between at the current and the next iteration. Our experiments confirm this technique to apply in reality system.