Exploring Methods to Enhance Appearance-Based Video Object Tracking using Dynamics Theory
Document typeMaster thesis
Rights accessOpen Access
The task of Video Object Tracking has for a long time received attention within the field of Computer Vision, and many different approaches have tried to tackle its challenges, being the ones based on appearance and motion some of the most popular ones. The main focus of this thesis is to fuse both strategies in order to exploit their strengths and overcome each other's flaws. To achieve this goal, we propose a unified framework that combines, in an online manner, an off-the-shelf single-object siamese tracker, which is modified to perform multi-object tracking and to provide more than one detection candidate, with a novel motion module. This module detects when the proposed target position is not dynamically consistent and, if that is the case, predicts an alternative which is used to choose the best among the rest of candidates. Our approach is evaluated on the challenging Similar Multi-Object Tracking (SMOT) dataset and achieves a relevant precision improvement of the 5% with respect to the baseline. We present an extension to the SMOT dataset, the eSMOT, including more sequences with complex dynamic scenarios, where the performance of our model is excellent, therefore we use its predictions to label the Ground Truth. Although there is still room for enhancement mainly regarding the efficiency of the approach, this work has served as a relevant proof of concept for the intuitions behind it and consequently, research in this direction will surely continue.
To be defined upon arrival.
DegreeMÀSTER UNIVERSITARI EN TECNOLOGIES AVANÇADES DE TELECOMUNICACIÓ (Pla 2019)
All rights reserved. This work is protected by the corresponding intellectual and industrial property rights. Without prejudice to any existing legal exemptions, reproduction, distribution, public communication or transformation of this work are prohibited without permission of the copyright holder