Sensor Fusion Using Dempster-Shafer Theory
Huadong Wu, Mel Siegel, Rainer Stiefelhagen, Jie Yang · 2002
Abstract- Context sensing for context-aware HCI challenges traditional sensor fusion methods with its requirements for (1) adaptability to a constantly changing sensor suite and (2) sensing quality commensurate with human perception. We build this paper on two IMTC2002 papers, where the Dempster-Shafer “theory of evidence ” was shown to be a practical approach to implementing the sensor fusion system architecture. The implementation example involved fusing video and audio sensors to find and track a meeting participant’s focus-of-attention. An extended Dempster-Shafer approach, incorporating weights representative of sensor precision, was newly suggested. In the present paper we examine the weighting mechanism in more detail; especially as the key point of this paper, we further extend the weighting idea by allowing the sensor-reliability-based weights to change over time. We will show that our novel idea – in a manner resembling Kalman filtering remnance effects that allow the weights to evolve in response to the evolution of dynamic factors – can improve sensor fusion accuracy as well as better handle the evolving environments in which the system operates.