Master of Science
This thesis presents the implementation and the quantitative and qualitative evaluation of three optical flow methods frequently used. We propose using a weight matrix of noise with three algorithms in computing optical flow. Our methods are implemented in traditional and hierarchical approaches of the Horn-Schunck and the Lucas-Kanade, and the Brox, respectively. Five different derivatives are used to compose the weight matrix, then estimating optical flow. We present quantitative results and give a qualitative evaluation based on various datasets. The tested data can be categorized into two groups, real imagery of nonrigidly moving scenes and realistic synthetic imagery. Our evaluation concentrates on several benchmarks such as angular error. The results showed that our proposed methods can produce a more precise flow field than that of original approaches.
Summary for Lay Audience
Optical flow is defined as the distribution of velocities of objects in images. It was widely used by robotics research in many areas such as object detection and tracking, movement detection, and robot navigation. Image noise, however, created when images captured, which challenges the accuracy of velocity estimates. This thesis presented an idea of using noise to improve velocity estimates produced from optical flow algorithms.
Liu, Wenyang, "An Investigation on Derivative and Model Noise for Optical Flow" (2019). Electronic Thesis and Dissertation Repository. 6300.