Intelligent Sensor Modeling and Data Fusion via Neural Network and Maximum Likelihood Estimation

Abstract

The major thrust of this paper is to develop a sensor model based on a probabilistic approach that could accurately provide information about individual sensor’s uncertainties and limitations. The sensor model aims to provide a most informative likelihood function that can be used to obtain a statistical and probabilistic estimate of uncertainties and errors due to some environmental parameters or parameters of any feature extraction algorithm used in estimation based on sensor’s outputs. This paper makes use of a neural network that has been trained with the help of a novel technique that obtains training signal from a maximum likelihood estimator. The proposed technique was applied to model stereo-vision sensors and Infra-Red (IR) proximity sensor, and information from these sensors were fused in a Bayesian framework to obtain a three-dimensional occupancy profile of objects in robotic workspace. The capability of the proposed technique in accurately obtaining three-dimensional occupancy profile and efficiently removing individual sensor uncertainties was demonstrated and validated via experiments carried out in the Robotics and Manufacturing Automation (RAMA) Laboratory at Duke University.