Multi-class classification and clustering based multi-object tracking

Nii Longdon Sowah, Qingbo Wu, Fanman Meng

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

Abstract

In this paper, we study the challenging problem of tracking multiple moving objects by a single camera. In our approach, a novel and efficient way to obtain an appearance and temporal based tracklet affinity model is proposed. We also propose to formulate the tracking problem as a classification task, where we classify tracklets into multi-classes, jointly across space and time. In our framework, we formulate our objective as a decision function with a Bayesian classifier constraint. We learn a model which seeks to maximize the decision function whilst minimizing our constraint. We estimate the probability density function of each class using a Gaussian probability density function, which reduces our misclassification loss. Our tracklet generation method minimizes the effect of missed detections and false positives. The proposed method emphasizes the effectiveness of multi-class SVMs (Support Vector Machines) in Multi-object tracking (MOT). Experimental results on three widely used Multi-Object Tracking datasets show that our method outperforms several state-of-the-art approaches in multi-object tracking.

Original languageEnglish
Title of host publicationProceedings of the World Congress on Engineering 2017, WCE 2017
EditorsDavid WL Hukins, A. M. Korsunsky, Len Gelman, S. I. Ao, Andrew Hunter
PublisherNewswood Limited
Pages480-484
Number of pages5
ISBN (Electronic)9789881404749
Publication statusPublished - 2017
Externally publishedYes
Event2017 World Congress on Engineering, WCE 2017 - London
Duration: 5 Jul 20177 Jul 2017

Publication series

NameLecture Notes in Engineering and Computer Science
Volume2229
ISSN (Print)2078-0958

Conference

Conference2017 World Congress on Engineering, WCE 2017
Country/TerritoryUnited Kingdom
CityLondon
Period5/07/177/07/17

Keywords

  • Classifier
  • Multi-class
  • Support vector machine
  • Tracklet

Fingerprint

Dive into the research topics of 'Multi-class classification and clustering based multi-object tracking'. Together they form a unique fingerprint.

Cite this