A new pedestrian navigation algorithm based on the stochastic cloning technique for Kalman filtering

In this paper we present a pedestrian navigation algorithm based on a Kalman filter that exploits relative position measurements provided by a step detection algorithm. In the development of pedestrian navigation systems we face certain challenges. In particular, GPS is often temporarily unavailable, especially in urban application scenarios. In addition, the sensor equipment has to be comfortably wearable by humans and is therefore constrained concerning its weight. For a broad application the cost of the overall sensor system is also a major concern. A common approach to meet these requirements is to exploit a low-cost IMU built with MEMS-technologies and additional sensors like a barometric altimeter and a magnetometer in order to estimate the position of the pedestrian when GPS is not available. Because the measurements from MEMS-IMUs are corrupted by substantial noise and biases, a direct integration of sensor readings provides a suitable position and orientation estimate only for a very limited period of time, typically a few seconds. One approach to alleviate these problems is to apply a technique called pedestrian dead reckoning where the orientation estimate is used to determine the direction of a step. The position estimate is obtained afterwards by concatenating the estimates of relative movement resulting from each step. In this two-stage approach to pedestrian navigation there is no estimate of the joint distribution over position and orientation available. Therefore, the correlations between them cannot be exploited during the estimation process. This aggravates sensor data fusion in the case that additional measurements from exteroceptive sensors or GPS measurements become available. We propose to use a technique known as "stochastic cloning" to enable direct integration of the relative position measurements arising from detected steps in a Kalman filter whose state vector comprises all relevant state variables. The main advantage of this approach is a correct treatment of the uncertainties arising from the delta measurements thus enabling accurate weighting of the state variables during sensor data fusion with exteroceptive sensors or GPS.