Tracker-Level Decision by Deep Reinforcement Learning for Robust Visual Tracking

Wenju Huang, Yuwei Wu*, Yunde Jia

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Citation (Scopus)

Abstract

In this paper, we formulate the multi-tracker tracking problem as a decision-making task and train an expert by the deep reinforcement learning (DRL) to select the best tracker. Specifically, the expert takes the response map of the tracker as input and outputs a scalar to indicate the reliability of the tracker. With the DRL, the expert can make full use of complementary information among base trackers. Furthermore, under the guidance of the deep expert, base trackers update themselves adaptively to capture the changes of object appearance and prevent corruption. The experimental results on public tracking benchmarks demonstrate that the proposed method outperforms the state-of-the-art methods.

Original languageEnglish
Title of host publicationImage and Graphics - 10th International Conference, ICIG 2019, Proceedings, Part 1
EditorsYao Zhao, Chunyu Lin, Nick Barnes, Baoquan Chen, Rüdiger Westermann, Xiangwei Kong
PublisherSpringer
Pages442-453
Number of pages12
ISBN (Print)9783030341190
DOIs
Publication statusPublished - 2019
Event10th International Conference on Image and Graphics, ICIG 2019 - Beijing, China
Duration: 23 Aug 201925 Aug 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11901 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference10th International Conference on Image and Graphics, ICIG 2019
Country/TerritoryChina
CityBeijing
Period23/08/1925/08/19

Keywords

  • Deep reinforcement learning
  • Tracker selection
  • Visual object tracking

Fingerprint

Dive into the research topics of 'Tracker-Level Decision by Deep Reinforcement Learning for Robust Visual Tracking'. Together they form a unique fingerprint.

Cite this