City Research Online

Abstract

Visual Odometry (VO) has its roots in the work of David Nister as an alternative to the problem encountered by the lunar rover. During NASA’s exploration of the moon using the lunar rover a method to track the trajectory of the agent was necessary to determine the location of the observations on the surface of the moon. In order to determine the location of the observations the rover was equipped with an encoder on the axis of its wheels, this would track the number of turns of the wheels which could then be integrated to determine the trajectory of the rover. Once combined with the initial conditions it was possible to estimate the position of the robot on the surface of the moon. In practice, this became problematic due to the nature of the moon’s surface, which led to the rover suffering from the phenomenon of wheel slip. This would lead to vast inaccuracy in the trajectory estimate over time.

Nister’s introduction of visual odometry relied upon the existence of key points in images, such as Harris corners, which could be traced between images and so projected into the three-dimensional world and projected back into the image plane given a stereo configuration. Which allowed for the estimation of the
rotation and translation undergone by the agent during this time.

With the progression of time, it became known that the visual odometry problem is a system of eight polynomial equations which can be reduced to four using the help of techniques such as Grobner basis. Various other facts also came to light with the progression of time such as the effects of flow decoupling and its use in calibrations.

The field also expanded to encompass many new sensors such as thermal cameras and Inertial Measurement Units (IMU). This led to the the development of new frontiers in the field such as data fusion and robustness analysis.

The introduction of the various sensor types has led to the adoption of various data fusion methods in the field. This includes many types of filters but also more sophisticated types of data fusion such as those employing Artificial Intelligence (AI). The introduction of these techniques has led to the development of many frameworks and packages for programming such solutions.

Whilst it is possible to utilise graph-based fusion techniques to combine all possible sensor solutions the approach does not facilitate the ability to use the unique properties of a sensor to compensate for the downfalls of another.

The advancement of AI has led to many unique innovations in the field of visual odometry. This is often in the form of a deep learning model with a different domain or the removal of a bias in the parameter estimation of the six degrees of freedom regression problem.

It is due to these changes that the current state of the field is a continuous expansion of the sensor configurations and data fusion methods employed. This is currently working hand in hand with the propagation of the field towards end-to-end deep learning solutions. A move away from the traditional approach to the field. The two aforementioned trends seem to have forgotten in large part the reconsolidation of the various sensors it has developed to encompass.

Publication Type:	Thesis (Doctoral)
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) T Technology > TK Electrical engineering. Electronics Nuclear engineering
Departments:	School of Science & Technology > Engineering > Electrical & Electronic Engineering School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads