Date of Award

9-25-2024

Publication Type

Thesis

Degree Name

M.A.Sc.

Department

Electrical and Computer Engineering

Keywords

Autonomous Vehicles;Computer Vision;Deep learning;Object Detection;Object Tracking

Supervisor

Ning Zhang

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Multi-object tracking (MOT) is a chore task in various applications, including au?tonomous driving and surveillance systems. Accurate and reliable MOT is essential for these systems to operate safely and efficiently, especially in dynamic and cluttered environments. Despite significant advancements in the field, challenges such as ob?stacles causing occlusions and background interference persist, often leading to false negatives, reduced accuracy, and reliability. Traditional MOT methods, such as the widely-used Deep SORT, face significant challenges in handling occlusions and minimizing false negatives. When objects are temporarily obscured, these methods can lose track of them, resulting in inconsistent object identities and increased false negatives. Addressing these challenges is vital for improving tracking performance in complex scenarios. This thesis enhances Deep SORT by incorporating memory management for oc?cluded items and utilizing the Sørensen-Dice coefficient for better similarity measure?ment. Our approach re-identifies occluded objects using motion features, enabling more robust tracking even when objects are temporarily obscured. Evaluated on the MOT16 dataset, our method significantly improves key perfor?mance metrics, achieving a MOTA of 61.84 and a recall of 69.8, compared to the baseline Deep SORT performance of 61.40 and 68.9, respectively. The contributions of this thesis to the field of MOT are significant, providing a reliable method for tracking objects in challenging situations. By addressing the limitations of existing methods, particularly in handling occlusions and reducing false negatives, this work paves the way for more reliable and accurate tracking systems in real-world applications, ultimately enhancing the performance and resilience of autonomous driving systems.

Share

COinS