A compact association of particle filtering and kernel based object tracking

  • Authors:
  • Anbang Yao;Xinggang Lin;Guijin Wang;Shan Yu

  • Affiliations:
  • Institute of Automation, Chinese Academy of Science, Beijing 100090, China;Department of Electronic Engineering, Tsinghua University, Beijing 100084, China;Department of Electronic Engineering, Tsinghua University, Beijing 100084, China;French National Institute for Research in Computer Science and Control, Inria, France

  • Venue:
  • Pattern Recognition
  • Year:
  • 2012

Quantified Score

Hi-index 0.01

Visualization

Abstract

Particle filtering (PF) and kernel based object tracking (KBOT) algorithms have shown their promises in a wide range of visual tracking contexts. This paper mainly addresses the association of PF and KBOT. Unlike other related association approaches which usually directly use KBOT to refine the position states of propagated particles for more accurate mode seeking, we elucidate the problem of what kind of particles is suitable for employing KBOT to refine their position states from a theoretical point of view. In accordance with the theoretical analysis, a two-stage solution is also proposed to resample propagated particles that are suitable for invoking KBOT from a computational perspective. The incremental Bhattacharyya dissimilarity (IBD) based stage is designed to consistently distinguish the particles located in the object region from the others placed in the background, while the matrix condition number based stage is formulated to further eliminate the particles positioned at the ill-posed conditions for running KBOT. Once the appropriate particles are obtained, constrained gradient based mean shift optimization enables us to efficiently refine the particles' position states. Besides, a state transition model embodying object-scale oriented information and prior motion cues is presented to adapt to fast movement scenarios. These ingredients lead to a new tracking algorithm. Experiments demonstrate that the proposed association approach is more robust to handle complex tracking conditions in comparison with related methods. Also, a limited number of particles are used in our association algorithm to maintain multiple hypotheses.