Abstract: Despite recent advances in multiple object tracking and pedestrian tracking, multiple-face tracking remains a challenging problem. In this work, the authors propose a framework to solve the problem in semi-online manner (the framework runs in real-time speed with two-second delay). The proposed framework consists of two stages: detection-tracking and tracklet-tracklet association. Detection-tracking stage is for creating short tracklets. Tracklet-tracklet association is for merging and assigning identifications to those tracklets. To the best of the authors’ knowledge, the authors make contributions in three aspects: 1) the authors adopt a principle often used in online approaches as a part of the framework and introduce a tracklet-tracklet association stage to leverage future information; 2) the authors propose a motion affinity metric to compare trajectories of two tracklets; 3) the authors propose an efficient way to employ deep features in comparing tracklets of faces. The authors achieved 78.7% precision plot AUC, 68.1% success plot AUC on MobiFace dataset (test set). On OTB dataset, the authors achieved 78.2% and 72.5% precision plot AUC, 51.9% and 43.9% success plot AUC on normal and difficult face subsets, respectively. The average speed was maintained at around 44 FPS. In comparison to the state-of-the-art methods, the proposed framework’s performance maintains high rankings in top 3 on two datasets while keeping the processing speed higher than the other methods in top 3.

Ngoc Q. Ly, Tan T. Nguyen, Tai C. Vong and Cuong V. Than, “The New High-Performance Face Tracking System based on Detection-Tracking and Tracklet-Tracklet Association in Semi-Online Mode” International Journal of Advanced Computer Science and Applications(IJACSA), 11(3), 2020. http://dx.doi.org/10.14569/IJACSA.2020.0110303

@article{Ly2020,
title = {The New High-Performance Face Tracking System based on Detection-Tracking and Tracklet-Tracklet Association in Semi-Online Mode},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2020.0110303},
url = {http://dx.doi.org/10.14569/IJACSA.2020.0110303},
year = {2020},
publisher = {The Science and Information Organization},
volume = {11},
number = {3},
author = {Ngoc Q. Ly and Tan T. Nguyen and Tai C. Vong and Cuong V. Than}
}