Loris Bazzani*, Marco Cristani*

1
Collaborative Particle Filters for Group Tracking

• Loris Bazzani, Marco Cristani, Vittorio
  Murino
• Speaker Diego Tosato
• Computer Science Department, University of
  Verona, Italy
• Istituto Italiano di Tecnologia (IIT), Genova,
  Italy

This research is founded by the EU-Project FP7
SAMURAI,grant FP7-SEC- 2007-01 No. 217899
2
Analysis of the problem (1)

• Multi-Target Tracking
• Estimate the trajectories of objects of interest,
  keeping their identification over the time
• Well-investigated problem
• State-of-the-art methods are very effective and
  efficient

• Multi-Group Tracking
• Estimate the trajectories of the groups of
  objects, keeping their identification over the
  time
• Not Well-investigated problem
• Few methods in the State of the art

3
Analysis of the problem (2)

• Why it is a hard task
• Methods for multi-target tracking fails
• Groups are highly structured entity
• Hard to model the complex dynamics
• Strong appearance variations over the time
• Intra- and inter-group occlusions phenomena
• What is a group?
• Motivation
• Highlighting social behaviors among individuals

4
Outline

• Overview of the proposed method
• Particle Filtering
• Multi-Object Tracking (MOT)
• Multi-Group Tracking (MGT)
• Collaborative Particle Filters (Co-PF)
• Results
• Conclusions

5
Overview of the proposed method

• Two separate particle filters
• Multi-object tracker (MOT) models each individual
  separately
• Multi-group tracker (MGT) focuses on groups as
  atomic entities
• Coupling of the two processes in a formal
  probabilistic framework

Co-PF Model
6
Particle Filtering for Target Tracking

• Recursively calculating the posterior
  distribution
• is defined by
• The dynamical model
• The observation model
• The first frame distribution
• Monte Carlo approximation by a set of weighted
  particles

7
Multi-Object Tracking

• Extension to Multi-target
• Hybrid Joint-Separable (HJS) Filter Lanz 2006
• Approximation to decompose the joint state space
  in single state spaces
• HJS is efficient and models the interactions
  among targets
• We just need to define
• Single-object dynamical and the single-object
  observation models

Lanz 2006 O. Lanz, Approximate bayesian
multibody tracking, IEEETPAMI, 28(9)14361449,
2006.
8
Multi-Group Tracking

• Use HJS filter
• State of the group Gaussian model
• Observation model
• Projection of the cylinder into the image
• Histogram-based feature as descriptor
• Dynamical model
• linear motion, perturbed by Gaussian
  noise
• Gaussian perturbation of its principal
  axes, i.e., by varying its eigenvalues and
  eigenvectors

9
Collaborative Particle Filters

• Inject the information collected by the MOT into
  the MGT
• Marginalization over the MOT state space
• After some approximations, we end up with
• It is a combination of MOT and MGT posteriors at
  time (t-1)

MOT posterior at time (t-1)
MGT posterior at time (t-1)
Linking probability
10
Collaborative Particle Filters

• The linking probability connect the MGT state
  space to the MOT state space
• Approximation through the Mixed-memory Markov
  Process
• Linking likelihood is decomposed in three
  components
• Appearance similarity distance between color
  histograms
• Dynamics consistency same direction between
  group and person
• Group membership spatial proximity between
  person and group

Linking likelihood
11
Results

• Compare Co-PF against MGT (without collaboration)
• An annotated dataset for group tracking does not
  exist
• Quantitative evaluation on a synthetic dataset
  emulating real scenarios

Kasturi et al 2009
ATA Average Tracking Accuracy MOTA Multiple
Object Tracking Accuracy MOTP Multiple Object
Tracking Precision FP False Positive MO
Multiple Objects FN False Negative TSR
Tracking Success Rate
Kasturi et al 2009 R Kasturi, D Goldgof, P
Soundararajan, V Manohar, J Garofolo,R Bowers, M
Boonstra, V Korzhova, and J Zhang,Framework for
performance evaluation of face, text, and
vehicledetection and tracking in video Data,
metrics, and protocol,IEEE TPAMI, 31(2)319336,
2009.
12
Results

• Qualitative evaluation on publicly available
  dataset

MGT
PETS 2009 dataset http//www.cvg.rdg.ac.uk/PETS200
9/a.html
13
Results

• Qualitative evaluation on publicly available
  dataset

Co-PF
PETS 2009 dataset http//www.cvg.rdg.ac.uk/PETS200
9/a.html
14
Results

• Qualitative evaluation on publicly available
  dataset

MGT
PETS 2009 dataset http//www.cvg.rdg.ac.uk/PETS200
9/a.html
15
Results

• Qualitative evaluation on publicly available
  dataset

Co-PF
PETS 2009 dataset http//www.cvg.rdg.ac.uk/PETS200
9/a.html
16
Conclusions

• A probabilistic, collaborative framework for
  multi-group tracking have been proposed
• Additional evidence on the individuals helps the
  group tracking in an effective way
• The results prove that the collaboration between
  trackers improve the performances
• Future directions
• Collaboration on the other direction (MGT
  MOT)
• Detection, split, and merge of the groups