Skip to main navigation Skip to search Skip to main content

Consistent online multi-object tracking with part-based deep network

  • Chuanzhi Xu
  • , Yue Zhou*
  • *Corresponding author for this work
  • Shanghai Jiao Tong University

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Multi-object tracking is still a challenge problem in complex and crowded scenarios. Mismatches will always happen when objects have similar appearance or are occluded with each other. In this paper, we appeal for more attention to the consistency of the trajectories and propose a part-based deep network which employs ROI pooling method to extract full and part-based features for the objects. An occlusion detector is proposed to predict the occlusion degree and guide the procedure of part-based feature fusion and appearance model update. In this way, the feature extraction speed of our tracker is faster, and the objects can be associated correctly even if they are partly occluded. Besides, we train the network based on siamese architecture to learn a dissimilarity metric between pairs of identities. Extensive experiments with multiple evaluation metrics show that our tracker can associate the objects consistently and gain a significant improvement in tracking accuracy.

Original languageEnglish
Title of host publicationPattern Recognition and Computer Vision - First Chinese Conference, PRCV 2018, Proceedings
EditorsCheng-Lin Liu, Tieniu Tan, Jie Zhou, Jian-Huang Lai, Xilin Chen, Nanning Zheng, Hongbin Zha
PublisherSpringer Verlag
Pages180-192
Number of pages13
ISBN (Print)9783030033347
DOIs
StatePublished - 2018
Externally publishedYes
Event1st Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2018 - Guangzhou, China
Duration: 23 Nov 2018 → 26 Nov 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11257 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference1st Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2018
Country/TerritoryChina
CityGuangzhou
Period23/11/18 → 26/11/18

Keywords

  • Consistent trajectories
  • Multi-object tracking
  • Occlusion detector
  • Part-based model

Fingerprint

Dive into the research topics of 'Consistent online multi-object tracking with part-based deep network'. Together they form a unique fingerprint.

Cite this