Multiple object tracking via deep learning techniques
Loading...
Date
2026-02-24
Authors
Υφαντή, Κωνσταντίνα
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Multi-Object Tracking (MOT) is a core problem in video understanding, with multiple practical applications. As a significantly challenging task, it has attracted significant attention from the computer vision community, with a wide range of approaches being continuously proposed. Most traditional techniques decompose the problem into separate, intuitive stages, such as object detection and data association. To address the limitations of this paradigm, a growing body of work proposes joint detection and tracking models. However, more recently, the introduction of the Transformer architecture and its success in natural language processing have motivated its adoption in multi-object tracking.
Beyond the inherent challenges of multi-object tracking and the complexities of each individual processing stage, the structure of the dataset plays a decisive role in a tracker’s generalization ability and overall performance. The shape, structure, and motion characteristics of objects can significantly affect tracking accuracy and introduce additional complexity. For this reason, this thesis presents a structured approach to studying deformity and motion complexity. By leveraging two different datasets, we aim to quantitatively analyze these characteristics at the dataset level.
Focusing on Transformer-based multi-object tracking methods, we analyze their core components and examine their effectiveness in the tracking task. Our study is based on a comparison between two benchmark datasets, DanceTrack and MOT17, which differ substantially in target deformity and motion complexity. Our objective is to investigate how attention mechanisms exploit available feature cues under varying dataset conditions and to assess whether appearance-based tracking models favor specific motion and deformation characteristics.
Description
Keywords
Παρακολούθηση πολλαπλών αντικειμένων, Υπολογιστική όραση, Νευρωνικά δίκτυα, Παραμόρφωση αντικειμένων