Return to Session D1 Next Abstract

Session D1: Robotic and Indoor Navigation

A Comparison of Particle Propagation and Weight Update Methods for Indoor Positioning Systems
Tanner Ray, Dan Pierce, and David Bevly, Auburn University
Location: Spyglass
Alternate Number 3

This work is a comparison of Particle Filters (PF) that utilize different particle propagation and weight update methods for indoor positioning systems. The particle filters fuse standalone PDR (Pedestrian Dead Reckoning) and a building map database to perform accurate pedestrian localization. A new type of weight update is also introduced which provides a more accurate and robust localization for indoor positioning. A detailed performance evaluation is presented with both simulated and experimental data.
In standalone inertial PDR systems, the position error grows with time due to the inertial measurement unit’s (IMU) internal sensor errors. Often external measurements from GPS, radio beacons, Wi-Fi, UWB radios etc. are used to restrict the error growth. The PF is a particularly useful estimation scheme for applying these external measurements. The presented work focuses on just using building maps and map matching techniques to constrain the standalone PDR system error. The map information is used to improve the position estimate by placing direct constraints on the state vector as a pedestrian’s movement is restricted by a building’s walls, floors, ceilings, and other features.
For indoor positioning systems, various methods exist in implementing the particle propagation and weight update. Different types of propagation and weight updates result from what is chosen to be the motion model, measurement likelihood, and the proposal distribution, and will impact the performance of the particle filter. The propagation is solely dependent on the motion model while the weight update depends on the motion model, measurement likelihood, and proposal distribution. The most popular methods form what is the conventional wall collision PF and is a type of bootstrap PF. The particle propagation is performed with the PDR as the motion model and the motion model is also used as the proposal distribution. The particles that ‘collide’ with a wall i.e. (the relative position vector between the a priori and a posteriori particle passes through a wall) are said to ‘die’ and given small or zero weights [1,2]. Other methods propose including some map information into just one part of the weight update, namely the proposal distribution, or can be a combination of the motion model, measurement likelihood, and proposal distribution [3,4]. These methods incorporate the map information by creating an angular PDF (Probability Density Function). The angular PDF weights the particles based on their proximity to the map. This creates a two-step process, first the particles are re-weighted using the angular PDF and then re-weighted again if the particle collides with the wall. Including map information into the weight update results in less particles colliding, which reduces particle degeneracy and sample impoverishment and thus the need to perform resampling.
This work proposes a new method in performing the weight update. The motivation for this new method is for when the map only constrains the particle cloud on a single side. The elimination of the particles on one side, shifts the mean of the particle cloud away from the wall. This produces a false estimate of the position and introduces drift into the solution. The proposal is to instead of weighting particles that collide to zero or a small number, to place these particles on the boundary of the wall and retain their weights. These particles are now no longer in violation and produce an estimate which reduces the drift and improves the navigation solution.
Simulated and experimental data were used to analyze the performance of the new and existing methods. Inertial measurements were taken with a MEMSENSE IMU that was mounted to the pedestrian’s foot. The IMU was mechanized with zero-velocity updates (ZUPTs) and a PDR solution generated using an Extended Kalman Filter(EKF). A 2-D building map was generated using precise measurements of the pertinent building characteristics. A virtual building was then created in an outdoor space, by mapping out lines on the ground to represent building walls and features. The outdoor setup allows for RTK GPS measurements to be taken, which provide a highly accurate external reference. This experimental setup also allows for control over the map’s geometry and the different methods can be tested in a wide range of scenarios.
The performance of the different methods and their combinations are evaluated with both simulated and experimental data. The PFs’ root-mean-square-error (RMSE) and resampling rates are compared and are produced with Monte Carlo simulations. The results of both experimental and simulated data show improved performance when using the new weight update. In particular, the presented method is less vulnerable to drifting away from a wall in an open area. With this superior position solution, the safety and effectiveness of navigating pedestrians are improved.
[1] Woodman, Oliver, and Robert Harle. "Pedestrian localisation for indoor environments." Proceedings of the 10th international conference on Ubiquitous computing. ACM, 2008.
[2] Davidson, Pavel, Jussi Collin, and Jarmo Takala. "Application of particle filters for indoor positioning using floor plans." Ubiquitous Positioning Indoor Navigation and Location Based Service (UPINLBS), 2010. IEEE, 2010.
[3] Nurminen, Henri, Matti Raitoharju, and Robert Piché. "An efficient indoor positioning particle filter using a floor-plan based proposal distribution." Information Fusion (FUSION), 2016 19th International Conference on. IEEE, 2016.
[4] Kaiser, Susanna, Mohammed Khider, and Patrick Robertson. "A human motion model based on maps for navigation systems." EURASIP Journal on Wireless Communications and Networking 2011.1 (2011): 60.



Return to Session D1 Next Abstract