Development of efficient object tracking algorithm using machine learning-based meta-learners

Abstract

Object tracking is one of the most complex tasks in the field of computer vision. Although there are many deep learning-based algorithms designed to tackle issues such as occlusion, low visibility, and scale changes, they typically require large amounts of data, powerful processing capabilities, and extensive training time. On the other hand, conventional algorithms like KCF, TLD, and BOOSTING might not be effective at tracking objects consistently. In order to bridge the existing gap, the primary focus of this study is to minimize the computing resources, training data, and training times required for object tracking. To accomplish this, a novel machine learning-based one-shot meta-learning approach has been proposed. The model combines traditional tracking algorithms, leveraging their predictions as inputs that are then refined through the machine learning model to improve overall performance. The evaluation of the proposed model on the VOT2015 and OTB50 datasets confirms its exceptional performance compared to existing deep learning-based approaches, offering a promising solution for the intricate challenges of object tracking in computer vision. By effectively combining traditional methods with machine learning techniques, this approach not only showcases impressive effectiveness but also underscores its potential to provide robust and resource-efficient object tracking solutions for real-world applications.