A New Pedestrian Navigation Algorithm Based On The Stochastic Cloning Technique For Kalman Filtering

摘要

In this paper we present a pedestrian navigation algorithmrnbased on a Kalman filter that exploits relative positionrnmeasurements provided by a step detection algorithm.rnIn the development of pedestrian navigation systems wernface certain challenges. In particular, GPS is oftenrntemporarily unavailable, especially in urban applicationrnscenarios. In addition, the sensor equipment has to berncomfortably wearable by humans and is thereforernconstrained concerning its weight. For a broad applicationrnthe cost of the overall sensor system is also a majorrnconcern.rnA common approach to meet these requirements is tornexploit a low-cost IMU built with MEMS-technologiesrnand additional sensors like a barometric altimeter and arnmagnetometer in order to estimate the position of thernpedestrian when GPS is not available. Because thernmeasurements from MEMS-IMUs are corrupted byrnsubstantial noise and biases, a direct integration of sensorrnreadings provides a suitable position and orientationrnestimate only for a very limited period of time, typically arnfew seconds.rnOne approach to alleviate these problems is to apply arntechnique called pedestrian dead reckoning where thernorientation estimate is used to determine the direction of arnstep. The position estimate is obtained afterwards byrnconcatenating the estimates of relative movementrnresulting from each step. In this two-stage approach tornpedestrian navigation there is no estimate of the jointrndistribution over position and orientation available.rnTherefore, the correlations between them cannot bernexploited during the estimation process. This aggravatesrnsensor data fusion in the case that additionalrnmeasurements from exteroceptive sensors or GPSrnmeasurements become available.rnWe propose to use a technique known as “stochasticrncloning” to enable direct integration of the relativernposition measurements arising from detected steps in arnKalman filter whose state vector comprises all relevantrnstate variables. The main advantage of this approach is arncorrect treatment of the uncertainties arising from therndelta measurements thus enabling accurate weighting ofrnthe state variables during sensor data fusion withrnexteroceptive sensors or GPS.