Location Sensor Fusion Mechanisms -- A Comparison

In order to detect the user's context or location, an application has to process position information from different sources (e.g. based on GPS sensors or cell IDs). A research field that traditionally deals with different position information is robotics. Based on e.g., odometers and ultrasonic distance measurement, a mobile robot computes the most probable location. Established approaches to compute a position from these sensors are Kalman and Particle Filters. Even though these approaches are widely used to process position sensor information, they often cannot be applied to mobile user scenarios due to significant differences: many positioning systems only have a limited coverage and availability and have a non-Gaussian measurement error distribution. Often, we receive a set of incomplete pieces of position knowledge such as: • I currently receive WLAN cell ABC; • checking my IP address, I know, I'm not at home; • one minute ago, I received GPS position xy, and I have a speed of 5 km/h as pedestrian. In the following we discuss important approaches to deal with such position information. All approaches have certain similarities: • As positions based on measurements have a limited precision, statements about positions (e.g. to exactly reside at position xy) cannot be mapped to true or false. Thus, all approaches model position knowledge with the help of probability densities that map every position to the probability to reside at this position. • All approaches deal with multiple measurements at the same time (the typical sensor fusion case) modelled by density operations. Multiple measurements are statistically represented by a multiplication of densities. • All approaches deal with movements, i.e. how two points in time and the respective densities are related. Movements are statistically represented by a convolution of densities. The approaches are very different, concerning how they represent probability densities and how they model multiple measurements and movements. The respective choices have significant influences on the usability in mobile user scenarios. In this paper, we discuss Kalman Filters, Particle Filters and Geometric Approaches considering the respective properties