Tracking More Than 100 Arbitrary Objects at 25 FPS Through Deep Learning
Lorenzo Vaquero, Víctor M. Brea, Manuel Mucientes
Pattern Recognition
SiamMT scales visual object tracking to dozens of arbitrary objects by reusing feature computations and introducing efficient crop-resize and pairwise similarity operators.
Most video analytics applications rely on object detectors to localize objects in frames. When real time is required, running the detector on every frame is usually not possible, so visual object trackers are instantiated between detector calls, but this does not scale with the number of objects. SiamMT is a deep-learning multiple visual object tracking solution that applies single-object tracking principles to multiple arbitrary objects in real time. It reuses feature computations, implements a crop-and-resize operator, and defines an efficient pairwise similarity operator. SiamMT scales to several dozen targets, reaching 25 FPS with 122 simultaneous objects for VGA videos or up to 100 simultaneous objects in HD720 video, while achieving leading performance against state-of-the-art trackers.
@article{vaquero2022tracking,
author = {Lorenzo Vaquero and
V{\'{\i}}ctor M. Brea and
Manuel Mucientes},
title = {Tracking More Than 100 Arbitrary Objects at 25 {FPS} Through
Deep Learning},
journal = {Pattern Recognition},
volume = {121},
pages = {108205},
year = {2022},
doi = {10.1016/j.patcog.2021.108205}
}