A Distributed Lifelong System for the Distributed Multi

A Distributed Lifelong System for the
Distributed Multi-Agent Multi-Object
Tracking Challenge
Fabio Previtali
15th October 2013
Rome, Italy
Outline
1
Problem statement
2
Multi-Clustered Particle Filtering
3
Experimental evaluation
4
BeeSafe project
5
Research plan and future directions
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
2 / 21
Multi-Object Detection and Tracking
A brief overview
Moving Object Detection
real-time extraction of moving objects from
sensors
Object Tracking [5]
continuous observation of the objects over time
to form persistent trajectories of the objects
5. Yilmaz et al., “Object tracking: A survey” in Journal ACM Computing Surveys (CSUR), 2006
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
3 / 21
A distributed lifelong system for the Distributed
Multi-Agent Multi-Object Tracking challenge
+
Ultimate Goal
Creating a distributed lifelong system able to continuously track the
objects in an environment using multiple sensors and that is able to learn
the behaviors of the tracked objects
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
4 / 21
Related work
Works that have significantly contributed to the improvement of one or
more aspects of multiple target tracking.
a synchronous algorithm based on the exchange of Gaussian Mixture
Models parameters that are evaluated using a distributed
Expectation-Maximization method [1]
a distributed approach that clusterizes the particle cloud in order to
detect the targets’ poses [4]
an asynchronous approach where the posterior probability is evaluated
using a gossiping protocol for data dissemination [2]
1. Gu et al., “Distributed particle filter for target tracking” in ICRA 2007
2. Oreshkin et al., “Asynchronous distributed particle filter via decentralized evaluation of Gaussian products”
in Information Fusion (FUSION), 2010
4. Wu et al., “Boosted Interactively Distributed Particle Filter for automatic multi-object tracking” in 15th IEEE
International Conference on Image Processing, 2008
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
5 / 21
Multi-Clustered Particle Filtering
A Distributed Multi-Agent Multi-Object Tracking method based on a
Multi-Clustered Particle Filter.
Input: a set of positions of the objects provided by a Multi-Object
detection system
Output: the estimated trajectories of the moving objects over time
It is not a general solution yet...
no target identity
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
6 / 21
Multi-Clustered Particle Filtering
Novelty
The novelties of the approach are:
a new clustering technique that keeps track of a variable unknown
number of objects
an improved robustness and reduced network overload by using
Gaussian Mixture Models
an asynchronous approach to improve the flexibility and the
robustness of the entire system
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
7 / 21
Multi-Clustered Particle Filtering
Estimation Process
Sensors
Agent a
z at
Particle Filter
ξ̃at
KClusterize
( λ at ,μ at , σ at )
1
1
1
ξ̃ tA
KClusterize
a
a
a−1
'
Distributed
Particle
Filter
(λt
a−1
,μ t
a +1
a +1
( λ t , μt
a −1
, σt
)
Network
( λ t ,μ t , σ t )
a +1
, σt )
a
( Λ t , Μ t , Σt )
'
'
'
( λ tA ,μ tA , σ tA )
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
8 / 21
Experimental evaluation
Application field
RoboCup Standard Platform League using the Aldebaran Nao humanoid
robots and considering the task of tracking the ball in the soccer field.
The Multi-Clustered Particle Filter method is the core of the PTracking
software library used for the experimental evaluation.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
9 / 21
Experimental evaluation
Objective
Demonstrate, using SimRobot [3] simulator, the robustness of the
PTracking software library by adding:
robot localization noise
perception noise
I
I
Increasing the duration of false perceptions
Increasing the number of false perceptions
We compared the performance, in terms of robustness, of our method with
respect to the algorithm developed by Wu et al. [4].
3. Röfer et al.: “B-human team report and code release 2011”. Technical report, 2011
4. Wu et al., “Boosted Interactively Distributed Particle Filter for automatic multi-object tracking” in 15th IEEE
International Conference on Image Processing, 2008
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
10 / 21
Experimental results
Nominal situation
We just introduced the robot localization noise.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
11 / 21
Experimental results
Increasing the duration of false perceptions
We introduced both the robot localization noise and we increased the
duration of false perceptions.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
12 / 21
Experimental results
Increasing the number of false perceptions
We introduced both the robot localization noise and we increased the
number of false perceptions.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
13 / 21
BeeSafe project
I am also working on the BeeSAFE project about “Learning for
Multi-Robot Task Allocation” for the Sistemi Software Integrati (SSI)
company.
Objective
Developing a system that is able to autonomously coordinate a team of
robots using learning techniques
The project involves two main activities:
1
SLAM Multi-Robot
2
Multi-Robot Task Allocation
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
14 / 21
Preliminary results
We successfully developed a system to learn the best path for a robot,
from a point A to a point B, in a simulative environment.
The high level actions defined using the
Petri Net Plans formalism.
Learning applied on the
choices points.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
15 / 21
Learning best combined navigational abilities
The experimental objective is to demonstrate that:
1
2
an optimal solution for a single robot can be different when
introducing multiple robots
implicit coordination among the robots in order to reach the best
combined solution
A best combined path for two robots in the same environment.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
16 / 21
Research plan
My research plan for the next year will focus on:
1
extending the designed method
I
I
2
implementing the ability of determining also the identities of the
tracked objects;
introducing a component for data association.
starting to learn basic behaviors of the tracked objects
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
17 / 21
Future directions
Tracking of individuals within a crowd using the PETS2009 dataset
Ground-truth Acquisition of Humanoid Soccer Robot Behavior
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
18 / 21
Ingegneria degli Algoritmi (type A): 30/30
CITEC Ph.D. Summer School (type B): seminar on 17th October with Prof. Iocchi to get a mark
Website of my seminars: www.cs.rug.nl/ds/Seminars/Program
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
19 / 21
Publications
“Distributed Multi-Agent Multi-Object Tracking with a
Multi-Clustered Particle Filtering” submitted to the 13th
International Conference on Autonomous Agents and Multiagent
Systems (AAMAS 2014)
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
20 / 21
Activities
RoboCup 2013, Eindhoven, The Netherlands
RoboCup German Open 2013, Magdeburg, Germany. Third place
RoboCup Iran Open, Tehran, Iran. First place
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
21 / 21
Dongbing Gu.
Distributed particle filter for target tracking.
In IEEE International Conference on Robotics and Automation, 2007, pages 3856–3861, 2007.
Boris N Oreshkin and Mark J Coates.
Asynchronous distributed particle filter via decentralized evaluation of gaussian products.
In 13th Conference on Information Fusion (FUSION), 2010, pages 1–8, 2010.
Thomas Röfer et al.
B-human team report and code release 2011.
Technical report, available online: www.b-human.de/downloads/bhuman11_coderelease.pdf.
Yi Wu, Xiaofeng Tong, Yimin Zhang, and Hanging Lu.
Boosted interactively distributed particle filter for automatic multi-object tracking.
In 15th IEEE International Conference on Image Processing, 2008. ICIP 2008, pages 1844–1847, 2008.
Alper Yilmaz, Omar Javed, and Mubarak Shah.
Object tracking: A survey.
Acm Computing Surveys (CSUR), 38(4):13, 2006.
A Distributed Lifelong System for the Distributed MAMOT Challenge
15th October 2013
21 / 21