Portland State University. Department of Electrical and Computer Engineering
Term of Graduation
Date of Publication
Master of Science (M.S.) in Electrical and Computer Engineering
Electrical and Computer Engineering
1 online resource (ix, 81 pages)
As machine learning and deep learning systems continue to find applications in science and engineering, the problem of providing these systems with high-quality data continues to increase in importance. Many of these systems utilize machine vision as their primary source of information, and in order to maximally leverage their abilities it is important to be able to provide them with high quality, accurate data. Unfortunately, many sets of tracking data extracted from video suffer from the problem of missing frames, which can arise from a multitude of causes depending on the system. These missing frames can result in confusion between object-track links, or even to the loss of tracking altogether.
This work is motivated by the specific problem of repairing flight tracking information in a pre-existing dataset involving large swarms of birds. These swarms can contain very large numbers of small birds, and the tracking information is derived from image segmentation of video, thus confusion between several objects in the data is a common and pervasive issue. The amount of data and the large number of tracked individuals necessitates efficient and trustworthy algorithms for interpolating between endpoints of known tracks.
We perform a literature review that discusses several approaches to track reconstruction and highlights the need for effective and accurate interpolation. We evaluate the effectiveness of a Kalman filter implementation on several tracks from such a dataset as well as synthetic tracks designed to simulate the flocking behavior of another species of interest. We also propose a novel interpolation algorithm, Stochastic Straight-Line Perturbation (SSLP), suitable for use on its own or as a pre-processing step for other algorithms. Using complete tracks, real and synthetic, with synthetically created gaps, we compare the accuracy of the Kalman filter to that of SSLP, as well as that of the Kalman filter operating on the same data pre-processed by SSLP. We find empirically that not only does SSLP outperform the Kalman filter on gap reconstruction in a number of scenarios, but it can also improve the performance of a Kalman filter, or potentially another tracking algorithm, when used as a pre-processing step. We also discuss performance trade-offs and potential limitations for its application.
For future work, we propose an alternate state-space model that may have some potential to improve upon the performance of the Kalman filter versus the model employed. We also discuss several possible strategies for further improving upon the performance of the SSLP algorithm both in terms of prediction accuracy and computational load, as well as several possible future applications.
In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).
Stamler, Zachary, "Methods for Object Tracking With Machine Vision" (2021). Dissertations and Theses. Paper 5635.