Inertial-Based Joint Mapping and

In this talk we will discuss reliable and high precision pedestrian positioning in GNSS denied environments. We will begin with a brief introduction into non-linear sequential Bayesian estimation principles and their application to sensor fusion. We shall present the p9roperties of pedestrian odometry (step measurement) based on foot mounted IMUs (Inertial Measurement Units) with zero velocity updates and the nature of the resulting error process. Next, we shall introduce an approach using sequential Monte-Carlo algorithms (“particle filters”) to stabilize the location estimate by the use of map information (e.g. a-priori known building layouts). Because such maps are not always available, we shall conclude with a presentation of purely inertial-based Simultaneous Localization and Mapping (SLAM) technologies and how to use them to learn the map based on pedestrian odometry data. Applications can range from offline map learning to real-time collaborative SLAM in unmapped environments. 1.0 INTRODUCTION Accurate long-term pedestrian localization is a particularly challenging task in GPS denied environments. While deep indoor reception is often possible with high sensitivity receivers, achieving meter-level accuracy is a different matter entirely. An approach that can be followed to address this problem is multisensor fusion, the principle being that sensors with different characteristics can augment each other. A particularly useful sensor in this context is the inertial measurement unit consisting of one or more accelerometers and/or gyros. The advantages of inertial measurements lie above all in its resilience and autonomy and its ability to measure short-term motion (changes). Recent research in the navigation community has focussed on pedestrian dead reckoning which uses inertial sensors to measure individual human steps. The result is a relative positioning system with reasonable short-term accuracy in two or three dimensions. To achieve long term accuracy, however, without any additional aiding such as GPS or wireless positioning one can resort to (known) building layouts that allow us to constrain the pedestrian’s estimated trajectory. The more a building constrains motion, the more accurate the resulting location accuracy. The main disadvantage is that an accurate building plan is required a-prior. This contribution is structured as follows. We begin with an exposition of sequential Bayesian estimation which is the theoretical and conceptual cornerstone of the localization and mapping techniques reviewed thereafter. The main flavours of pedestrian dead reckoning are then explained and this serves as the basis for map-aided approaches. The main contribution of the paper is to motivate and introduce the concept of odometry based simultaneous localization and mapping (SLAM) for pedestrians. The objective of SLAM – a principle developed by the robotics industryis to perform a joint estimate of maps and location based on sensor measurements, which means that maps (or building plans, in our case) are not needed a-priori. In robotics, sensors used in SLAM comprise devices such as laser scanners or cameras in addition to odometry (motion change) information from the robot. The main variant of SLAM for humans Inertial-Based Joint Mapping and Positioning for Pedestrian Navigation 9 2 RTO-EN-SET-116(2011) (“FootSLAM”) discussed here, however, needs no sensors apart for the inertial-based human step measurement. Before sketching the mathematical foundations we shall present some important application scenarios of FootSLAM. This is followed by an example of the attainable performance accuracy. Finally, we present an augmentation of FootSLAM – “PlaceSLAM” which requires active participation of the pedestrian to improve mapping and location accuracy. 2.0 NON-LINEAR SEQUENTIAL BAYESIAN ESTIMATION 2.1 Principles If we are interested in finding out as much as possible about the state of a system we need a measurement device that provides an output (“measurement”) that is somehow influenced by the state of interest. In our application domain the state of interest might be the position of a pedestrian and the measurement device a GPS receiver, whose output hopefully depends on the actual position of the bearer. In real-world physical systems the state cannot change arbitrarily fast, due to some form of inertia. Hence, knowledge about a system’s previous state tells us something about the system’s current state. In order to obtain the best possible estimate of some state variable x at a time instant k, we should not only consider the available measurements, but also the past evolution of this state variable. If all information that we consider relevant for the future evolution of the state is represented in that state variable, say the current position and velocity for our location determination problem, we can model the evolution as a first order Hidden Markov Model as depicted in Figure 1. “Likelihood” p(zk+1| xk+1) zk-1