Stereo-Vision-Aided Inertial Navigation

Reliable information about position and attitude is an essential requirement for many applications. The work expounded in this paper aims at a tight integration of low-cost inertial navigation and stereo vision to obtain this information. The method I present here is based on passive measurements and does not rely on external referencing. Thus, it provides a navigation solution for unknown indoor and outdoor environments. Special attention is paid to a stereo vision-based system, capable of providing egomotion measurements with six degrees of freedom. Natural landmarks are extracted and tracked in consecutive image pairs aided by inertial measurements to constrain the correspondence problem. This effectively reduces the computational effort and avoids uncertainties that stem from mismatches, resulting in a robust tracking algorithm that runs in real time. In turn, the extracted egomotion is used to constrain the inertial sensor drift. In addition, the measured gravity serves as a vertical reference, stabilizing the navigation solution. Based on dead reckoning, inertial navigation is widely used and has been studied in almost every aspect. To correct the inertial sensor errors, these systems are periodically provided with external data, e.g. a Global Positioning System (GPS) signal. The reliable short term properties of inertial data are complemented by the long-term stability of the external measurement. Although such methods do work very well, a similar solution is needed for navigating in difficult environments with erroneous or no external reference whatsoever. In such situations, using independent measurement systems like barometers, odometers, or vision-based systems is especially advantageous. Hence, I present an approach for a heterogeneous multi-sensor system that involves both hardware and software, wherein aspects like synchronization, registration, and calibration of the sensor system are considered. As the optical system is of major importance, I’ve developed a new method that provides fast and reliable camera calibration. Herein, I also present my extensive analysis of possible error sources throughout the system. The result of this integration of stereo vision and inertial navigation is then proven in various pedestrian navigation tasks.