В библиотеку
Sensor Fusion Using Dempster-Shafer Theory
Huadong Wu, Mel Siegel, Stiefelhagen Rainer, Jie Yang
Abstract
Context-sensing for context-aware HCI challenges the traditional sensor fusion methods with dynamic sensor configuration and measurement requirements commensurate with human perception. The Dempster-Shafer theory of evidence has uncertainty management and inference mechanisms analogous to our human reasoning process. Our Sensor Fusion for Context-aware Computing Project aims to build a generalizable sensor fusion architecture in a systematic way. This naturally leads us to choose the Dempster-Shafer approach as our first sensor fusion implementation algorithm. This paper discusses the relationship between Dempster-Shafer theory and the classical Bayesian method, describes our sensor fusion research work using Dempster-Shafer theory in comparison with the weighted sum of probability method. The experimental approach is to track a user’s focus of attention from multiple cues. Our experiments show promising, thought-provoking results encouraging further research.
Introduction
The ultimate goal of context-aware computing is to have computers understand our real physical world. We envisage “smart environments”, where human-computer-interactions (HCI) feel natural, as if we were communicating with human assistants or service personnel. This seems an impossible task today, as it faces a two-fold challenge (1) how properly to represent our colorful world, with its abstract concepts and unpredictable human feelings, in a computer understandable way, and (2) how to design and deploy sensors that can sense all the clues and content and map them into the context representation.
Researchers in the context-aware computing community are now challenging this seemingly insurmountable task, starting with simple context information, such as a user’s location. The argument we offer is that (1) although a user’s situational context may include complex information, we can always try to decompose the hard-to-explain complex information contents into simpler and less abstract information pieces; (2) among various relevant information sources about a user’s activities and intentions and the current environment, some pieces of information (such as location and user identification) have already been demonstrated (see background research of [1][2]) to be both useful and not too difficult to implement.
A user’s context content may include various aspects of relevant information; meanwhile, different sensors have different measurement targets, different resolutions and accuracies, and different data rates and formats. Thus, the mapping from sensors’ output to context information can be extremely complicated. Generalized solution do not exist, and systematic research to asymptotically approach it has hardly begun yet. Our research aims to push forward in this direction. Our current work mainly deals with intermediate and higher level symbolic or modality fusion vs. lower-level signal processing.
In this paper, we deal with simplified situations where we assume that the context information can be represented by discrete symbols or numbers, and the mapping from sensor output data to the context representation data structure is well defined. Specifically, our starting point is that the sensors we will use can generate context information fragments that comply with our context representation specifications. They report both context information and the corresponding confidence estimations. However, sensor outputs often have overlaps and conflicts, sensors are highly distributed, and sensor configuration is very dynamic (sensors come and go, sensors’ performance varies all the time). Our goal is to build a system framework that manages information overlap and resolves conflicts. This system provides generalizable architectural support that facilitates sensor fusion, i.e., the sensor-to-context mapping process.
Our approach to achieve this goal is to use layered and modularized system design, in which the sensed context information is separated from the sensors’ realization, and the sensor fusion process is analogous to the human perception and reasoning processes. Using the Dempster- Shafer theory of evidence algorithm as our baseline sensor fusion approach reflects this analogy.
The full text can be downloaded from: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.15.3941&rep=rep1&type=pdf