Publication

Abstract : The long-term sub-track of visual object tracking challenge comprises of some of the most challenging scenarios like occlusion and target disappearance and reappearance. To this end, many deep learning solutions with multiple levels of detection have been proposed. Most of these solutions tend to re-identify a wrong target during the occlusion or disappearance as they start looking for the target in the entire frame. Instead, through this work, we intend to prove that predicting a probable search region for the target by understanding its trajectory and searching for a target in it will help in reducing the misidentifications and also aid in the increase of IoU. For this, we have utilized the trajectory modeling capabilities of the Kalman filter. With this proof of concept work, we achieved an average improvement of 37.37% in IoU in the sequences where we overperformed MBMD.