Knowledge of your own and other's positions are frequently a prerequisite for acting, leading others, and interacting in and with the environment; to retrieve relevant information and to process and interpret it; and to understand, compile, and learn from observations of the surrounding and its dynamics. This holds for humans as well as for machines and systems made for supporting and controlling them. Consequently, systems which automatically provide position information of peoples are of interest and the larger subject area of this thesis. Position can be obtained from well-known infrastructure based systems such as GPS. However, these systems carry obvious drawbacks in their infrastructure dependence which gives them limited coverage and system robustness. By observing our own ability to localize ourselves, it is obvious that localization without infrastructure at with a better (relative) accuracy is achievable. The development over the last decades of sensor and processing hardware and statistical methods have started to make such localization possible. This thesis specifically concerns systems and statistical methods for infrastructure-free localization. The this primarily deals with statistical methods but also describe hardware in terms of high-level system designs.
For many critical applications such as positioning of emergency responders, dismounted soldiers, and security personnel, it is unsuitable for the positioning system to be dependent on infrastructure or prior knowledge about the environment. Consequently, this thesis deals with systems and methods for infrastructure-free and prior-knowledge-free pedestrian localization. The thesis is specifically concerned with statistical methods but will also cover hardware in terms of high-level system designs. The thesis is composed of an introduction followed by a collection of papers which are divided into two parts, each concerning a separate problem area. The introduction motivates and describes the localization problem in general terms and gives a coherent guide to the articles.
The first group of articles together describes an infrastructure-free system for tactical localization of small units of agents. The physical implementation of the localization system carries the name TOR (Tacitcal lOcatoR) and have been tested on fire fighters during realistic smoke diving exercises. This system primarily depends on pedestrian dead-reckoning based on foot-mounted inertial navigation and inter-agent radio ranging. The core parts of the system which are dealt with are: foot-mounted inertial navigation units which provides dead reckoning of individual agents, system structure and estimation algorithms which, based on the dead reckoning and inter-agent ranging, provides estimates of the agent positions, initialization algorithms for the estimation, and a user interface which exploits voice radio communication and 3D-audio to let the agents hear where they have each other.
The second group of articles concerns low-level processing for extraction of spatial information of camera images (video), a prevailing infrastructure-free data source for relating an agent's position to the environment. These articles are focused on formalization and fast implementations of fundamental processing steps. An implementation of scale-space only relying on integer signal representation of image data and simple arithmetic operations is presented. Further, a unifying theory of feature point orientation assignment is derived and a novel method for the same is presented. Thereafter, the small but frequently occuring processing step in which image gradient samples are binned based on their argument, is treated and three fast solutions with varying properties are suggested. Finally, a localization system based on inertial navigation aided by imagery data is presented.
Stockholm: KTH Royal Institute of Technology, 2013. , viii, 37 p.