The hybridization of GNSS with IMU, odometer and magnetometer, and the geometric processing of images from a video camera is a promising approach that can possibly maintain accuracy at acceptable levels even when IMU drags the position error with an increasing drift. To our knowledge, this GNSS/Video/IMU combination has already been exploited by various authors with good results, even though in situations where GNSS RTK-PPP or high-grade IMU is impracticable, the resolution of the accumulated drift is sometimes difficult, but complete characterization of the error is not provided as a general result but from collected datasets, and the main unknown at this stage if the performance of the video-extracted geometric observations.

The featured complementary sensor in this combination is the video sensor, which is expected to bring user trajectory information from the image sequence, therefore in a manner which is not dependent on reception of RF signals, and which is not a pure dead reckoning building up errors. However, scene environment, scene luminance, camera technology and tuning, user dynamics, geometric processing of images will strongly influence the quality of the trajectory information brought by the camera to the system. The geometric processing of the video can be accomplished via line-following, visual odometry, perspective-and-point, and SLAM. The favoured technique at this stage is SLAM, which allows extracting visual landmarks in the observed scene for the joint estimation of the position of these reference points and the position and orientation of the camera. A full state of the art review will complete the landscape of applicable techniques. SLAM has many setting parameters for initialization, feature detection, feature description, feature matching, feature tracking, key frame identification, local and global optimization, and its camera pose estimation error performance is known to vary upon these parameters in size and in correlation time. In addition, camera tuning such as frame rate, field of view, shutter, resolution, focus, gain, can influence the resulting errors.

To our knowledge, integrity bounding of the position error of such a combination, i.e. reliably predicted 95%-accuracy, integrity, continuity and availability performance of such a combination is not available in the literature, except in some well bounded cases. Several long steps are needed to reach that integrity goal, including development of performance requirements for applications, identification of nominal measurement error models including correlation time, identification of failure modes, development of monitors with associated detection thresholds, prediction of monitors missed detection performance and minimum detectable biases, determination of accuracy and integrity bounds on position error, prediction of availability and continuity performance. This proposed thesis is an additional step towards the goal of integrity bounding, but is not reaching that overall goal of inte.

To derive the measurement error models, existing or new data collections will be used to infer the shape, the parameters of the models and in particular the description and the limitations of applicability for the environment, but it is felt that the performance prediction cannot rely only on data collection which has inherent limitation in extension of cases, plus inherent unknowns about the truth of the values of several parameters like the scene and the conditions of visibility.

The purpose of this thesis is to develop advanced models and techniques for the optimal fusion of GNSS, video, IMU, odometer and magnetometer data in order to obtain an accurate and robust hybrid positioning system in constrained environments and to reliably predict a 95% position error bound in a representative set of conditions. As a by-product, it is also proposed to analyze further the velocity error models and distributions.

Expected original contributions:

1.The development of nominal error models and models of video measurement geometric anomalies resulting from video processing as a function of descriptive variables (environment, visibility, travelled distance, …). These models will be inferred from existing or collected real data, and will also be checked against real data.

2.The development of a data fusion algorithm tuned on the basis of these models. The starting point is the existing combination algorithms

3.The development of a GNSS/Video/IMU positioning simulator reliably implementing the developed models and the combination algorithm.

4.The analysis and the prediction of a reliable 95% accuracy performance bound based on the descriptive variables for a variety of defined situations, and the comparison with other state-of-the-art approaches.

5.As a by-product, the analysis of the velocity error models and distributions.