Abstract

Deep Learning allows for great advancements in computer vision research and development. An area that is garnering attention is single object tracking and multi-object tracking. Object tracking continues to progress vastly in terms of detection and building re-identification features, but more effort needs to be dedicated to data association. In this thesis, the goal is to use a graph neural network to combine the information from both the bounding box interaction as well as the appearance feature information in a single association chain. This work is designed to explore the usage of graph neural networks and their message passing abilities during tracking to come up with stronger data associations. This thesis combines all steps from detection through association using state of the art methods along with novel re-identification applications. The metrics used to determine success are Multi-Object Tracking Accuracy (MOTA), Multi-Object Tracking Precision (MOTP), ID Switching (IDs), Mostly Tracked, and Mostly Lost. Within this work, the combination of multiple appearance feature vectors to create a stronger single feature vector is explored to improve accuracy. Different types of data augmentations such as random erase and random noise are explored and their results are examined for effectiveness during tracking. A unique application of triplet loss is also implemented to improve overall network performance as well. Throughout testing, baseline models have been improved upon and each successive improvement is added to the final model output. Each of the improvements results in the sacrifice of some performance metrics but the overall benefits outweigh the costs. The datasets used during this thesis are the UAVDT Benchmark and the MOT Challenge Dataset. These datasets cover aerial-based vehicle tracking and pedestrian tracking. The UAVDT Benchmark and MOT Challenge dataset feature crowded scenery as well as substantial object overlap. This thesis demonstrates the increased matching capabilities of a graph network when paired with a robust and accurate object detector as well as an improved set of appearance feature vectors.

Library of Congress Subject Headings

Computer vision; Optical pattern recognition; Neural networks (Computer science)

Publication Date

5-11-2021

Document Type

Thesis

Student Type

Graduate

Degree Name

Imaging Science (MS)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Andreas Savakis

Advisor/Committee Member

Guoyu Lu

Advisor/Committee Member

Anthony Vodacek

Campus

RIT – Main Campus

Plan Codes

IMGS-MS

Download

Share

COinS