Loading...

Improving the Performance of Distributed Fusion for PHD Filter in Multi-Object Tracking

Khazaei, Mohammad | 2014

517 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 46084 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Jamzad, Mansour
  7. Abstract:
  8. The Gaussian mixture (cardinalized) probability hypothesis density (GM-(C)PHD) filter is a closed form approximation of multi-target Bayes filter which can overcome most of multi-target tracking problems. Limited field of view, decreasing cost of cameras and its advances induce us to use large-scale camera networks. Increasing the size of camera networks make centralized networks practically inefficient. On the other hand, scalability, simplicity and low data transmission cost has made distributed networks a good replacement for centralized networks. However, data fusion in distributed network is sub-optimal due to unavailable cross-correlation.Among data fusion algorithms which deal with unavailable cross-correlation, Covariance intersection (CI) has good performance, but before using CI, we need to optimize fusion weights. Kullback–Leibler divergence (KLD) and Rényi divergence (RD)arecommon measuresof the difference between two probability distributions. They widely used in many applications, but they don’t have any closed form solution for Gaussian mixture models (GMM). In this thesis, we evaluatethe use of difference approximations to KLD and RD of two GMMs for optimizing CIfusion weights in GM-PHD/CPHD filter on a simulated scenario. Also using of Fast Covariance Intersection (FCI) with those approximations is evaluated and compared to above methods.Then, the proposed method for fusion of GM-CPHD in partially overlapping camera is evaluated on a simulated scenario. Finally, a multi-human tracking framework using GM-PHD filterwith track-after-detect scheme for distributed camera network is proposed. The proposed method is evaluated on public PETS2009 dataset.The result measured in Correct Tracking Percentage(CTP) metric shows better performance compared with one of the most recent related works in the evaluated datasets
  9. Keywords:
  10. Distributed Tracking ; Data Fusion ; Multitarget Tracking ; Probability Hypothesis Density (PHD)Filter

 Digital Object List

 Bookmark

No TOC