Philippine Journal of Science
150 (3): 607-618, June 2021
ISSN 0031 – 7683
Date Received: 23 Jul 2020

Multiple-object Tracking Based on Movement
Direction Assumption and Potential Re-appearance
Position with Object Flow Visualizations

Yun Sup Lee* and Joel Ilao

Computer Technology Department, De La Salle University
Manila, Metro Manila 1004 Philippines

*Corresponding author: yun_sup_lee@dlsu.edu.ph

ABSTRACT

This paper proposes a multiple-object tracking (MOT) approach that adopts the tracking-bydetection strategy, which is composed of two steps: detection and tracking. This study aims to improve the tracking module by integrating a kernelized correlation filter (KCF). The filter is modified to be scale-adaptive to manage the varying size of the target object. Moreover, the tracking module’s data linking is performed in two steps, i.e. data association and reidentification (ReID). Here, two novel assumptions – namely, the movement direction assumption and potential re-appearance position – are incorporated. The first assumption considers the target object’s expected movement direction, while the second assumption postulates the possible position that a missing object may re-appear in. Using selected videos in 2DMOT2015 and PNNL Parking Lot benchmarking datasets, this study demonstrates that the proposed method outperforms the baseline model. In addition to MOT, this study further introduces a simple yet effective technique called trajectory accumulation to visualize the target objects’ flow of movement using the trajectories generated by the proposed tracker.