[Back]

G. Retscher:
"Multi-sensor Fusion Using a Kalman Filter and Knowledge-based Systems";
Talk: ION GNSS 2009 Conference, Savannah, Georgia, USA; 2009-09-22 - 2009-09-25; in: "ION GNSS 2009 Conference", (2009), 8 pages.

Navigation applications require the integration of different sensors to be able to perform continuous position determination of the user. Apart from GNSS, alternative technologies such as Wi-Fi, UWB, RFID and aiding sensors such as MEMS accelerometers and gyros as well as magnetometers, etc. can nowadays be integrated in a multi-sensor system. Sensor fusion for multi-sensor navigation systems is usually based on a Kalman filter. As a central element of the location determination module the Kalman filter provides an estimate of the current state of the user in real-time. Observation errors or an inadequate dynamic model can cause unsatisfactory filter results (e.g. large deviations of the current location from the truth) which can lead to a divergence of the filter in the worst case. Using a knowledge-based pre-analysis of the observations and an appropriate stochastic model of the Kalman filter such cases can be prevented. For instance, gross errors and outliers in the observations can be detected and eliminated before the observations are introduced into the filter. The knowledge-based component can handle the following tasks:
- Pre-filtering of all observations of the different location sensors,
- selection of an appropriate dynamic filter model,
- calibration of the multi-sensor system, and
- management of the sensors.
The pre-filtering of the sensor observations runs parallel to the data acquisition and provides suitable observations for the following central Kalman filter. The knowledge-based component estimates from the obtained velocity and acceleration measurements (e.g. from a low-cost IMU) the current userīs dynamics (e.g., standing, walking, running, etc.) and from the observations of the heading sensors in the last epochs the geometry elements of the userīs trajectory (e.g., if it is a straight line or arc). Depending on this estimation, the dynamic filter model is selected. The system calibration consists of the initialization of the multi-sensor system in the beginning and necessary calibrations during operation to reduce the sensor drifts. The management of the sensors handles practical issues during operation, e.g. the activation of power safe modes for sensors which are currently not in use (e.g. deactivation of the GNSS sensor if the user moves indoors).
In this paper, this approach for multi-sensor fusion is discussed in detail. As a use case, the navigation of a pedestrian in a combination of an urban outdoor environment and an indoor environment is investigated. Field test results show that a pedestrian user can be located with a positioning accuracy in the range of 2 to 5 meters using such a multi-sensor approach.