Object-Aware Anchor-Free Video Object Tracking using Attention Mechanism and Target Dynamics
Ali Abbasi Boroujeni, Rahil Mahdian Toroghi, Hassan Zareian · 2024
In this paper, we have proposed a novel method for video object tracking that involves an object-aware system in order to follow the target object dynamically. An extra processing path has been considered that incorporates a transformer model and covers the temporal variation of the target object and provides the required information for the main tracking path. The object localization architecture is chosen to be anchor-free in order to avoid the request for predefined proposal regions of the target object. This part of the system directly estimates the boundaries of a target-contained region. By implementing the proposed structure on GOT-10K dataset, and comparing the proposed architecture with some state-of-the-art methods the results outperformed these methods using the Average Overlap (AO) and Success Rate (SR) measures, significantly.