Neural network and kalman filter use for improvement of inertial distance determination

Abstract

Appropriate distance estimation is very important in different applications, e. g. in navigation or developing natural interfaces forman-machine interaction. Article refers to this problem and presents two approaches in improving estimation of the distance. The distance is computed on the base of linear acceleration. The acceleration data is captured by an inertial sensor mounted on moving object. The first approach uses Kalman filter and appropriate preprocessing steps to denoise measured acceleration. This method improves the distance estimation in noticeable manner but is not optimal because of time growing errors. These errors results come from the imperfection of the accelerometer and double integration of acceleration data during computational step. The second approach improves the estimation accuracy by using a neural network. The neural network estimates position of moving object on the base of statistical properties of the acceleration signal. Both of mentioned approaches were compared and the results are described in this article. Theoretical contemplation was confirmed by practical verification which results are also presented. Conducted research show that these two approaches can be combined for an optimal problem solution.