Author: Gayeon Lee
Lee, Gayeon, 2023 Vision-Based Real-Time Velocity Estimation, Flinders University, College of Science and Engineering
Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.
For autonomous system, it is often useful to know the position and velocity of nearby objects for tasks such as obstacle avoidance or object tracking. The hidden states of an object, such as velocity, can be estimated by knowing the mathematical model that describes the relationship between the measurable state and the hidden state. Hence, both position and velocity of the object can be found by object position measurement. For autonomous Micro-Aeria-Vehicle (MAV), various research investigated in using camera for real-time object state estimation as it is a low-cost and low-complexity alternative of ranging sensors. However, little literature is available when the scope is narrowed down to a moving object and moving MAV, despite the position of ground object can be obtained with the camera on MAV. This project aims to develop real-time object velocity estimation system, only using strapdown camera for the exteroceptive sensor, to be integrated onto a MAV (or other aerial system). The project assumes the object is a ground vehicle, on a flat surface with known altitude, moving at a constant acceleration. The system must be robust to measurement noise for practical application and must be applicable for object at a long-range (i.e. 100 m+). The ground vehicle (for the object) and MAV was simulated at maximum of 30 and 50 m/s of speed. Considering the reviewed literatures, this project initially proposed to use the pin-hole camera model, with Extended Kalman Filter (EKF) as a baseline, and to fuse with optical flow model for inertial state-estimation by Unscented Kalman Filter (UKF). The baseline EKF was developed and tested on a sample dataset, then in depth analysis of EKF was done to assess its suitability for velocity estimation using a strapdown monocular camera. The dataset for system analysis was extracted which consisted of scenarios with stationary vehicle, vehicle in linear and angular motion in a combination of different MAV motion. The baseline system was able to achieve all project aim in the tested scenarios except achieving robust performance under measurement noise; it was found the system has significantly high sensitivity to noise in MAV orientation. It was assumed that the system sensitivity is caused from the measurement model. The rotation matrices, that contains MAV orientation and camera parameters, in the measurement model is highly non-linear. EKF uses linear approximation of model which is not as effective to highly non-linear model where UKF performs better in such cases. As the future work, including system application to UKF, a list of tasks was identified to improve the system and further analyse the system performance.
Keywords: Extended Kalman Filter, Vision-based state estimation, Flat-surface assumption, Monocular camera, Pinhole camera model
Subject: Engineering thesis
Thesis type: Masters
Completed: 2023
School: College of Science and Engineering
Supervisor: Dr Karl Sammut