GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking...

20
1 www.hm.pjwstk.edu.pl 1 GPU – Accelerated Tracking of the Motion of 3D Articulated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad Wojciechowski Warszawa 2010-12-28

Transcript of GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking...

Page 1: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

1www.hm.pjwstk.edu.pl

1

GPU – Accelerated Tracking of the Motion of 3D Articulated Figure

Tomasz KrzeszowskiBogdan KwolekKonrad Wojciechowski

Warszawa 2010-12-28

Page 2: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

2www.hm.pjwstk.edu.pl

• The problem

• CUDA• Parallel PSO

• Tracking framework• Demonstration of tracking effectiveness

• Experimental results

Plan

Page 3: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

3www.hm.pjwstk.edu.pl

• Articulated human body motion tracking• High dimensional search space• Currently there is no on-line 3D model-based human tracking algorithm

The problem

Page 4: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

4www.hm.pjwstk.edu.pl

• Human body motion tracking is done via PSO algorithm• GPU speed-up is realized by CUDA parallel architecture

The contribution

Page 5: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

5www.hm.pjwstk.edu.pl

CPU vs. GPU

1. www.intel.com

2. www.nvidia.com

6.4

≈ 0.05

≤ 8

SISD

CPU (Intel Xeon) 1

141.7bandwidth (GB/s)

≈ 1TFLOPS

≥ 100core number

SIMDarchitecture type

GPU (GTX 280) 2

Page 6: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

6www.hm.pjwstk.edu.pl

• Compute Unified Device Architecture• Scalable to 100’s of cores and 1000’s of parallel threads

• Small set of extensions to C language• Let programmers to focus on parallel algorithms• Supports heterogeneous systems (i.e., CPU+GPU)

CUDA

Page 7: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

7www.hm.pjwstk.edu.pl

PSO• Particle Swarm Optimization

• Population-based evolutionary algorithm forfinding best solution for the problem

Page 8: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

8www.hm.pjwstk.edu.pl

PSO

ii fwfwfwxf +++= L2211)(

)()( )(best

)()(best

)()( iiiii xgrcxprcvv −+−+← 2211ω

Page 9: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

9www.hm.pjwstk.edu.pl

SynchronousPSO - algorithm

max_iter

vxx

xgrcxprcvv

pggfpf

xppfxf

f

xfgxp

vx

xg

iii

iiiii

ii

iiii

i

x

ii

ii

i

i

n evalutatiofunction ofnumber Until.

:particle each ofposition ate Upd.

)()(

:particle each of velocity ate Upd8.

valuebest global the update ),()( If .

valuebest local the update ),()( If 6.

particles all for)( function Evaluate 5.

Repeat.

)(minarg,.

,Initialize.

)(PSOFunction.

)()()(

)(best

)()(best

)()(

)(bestbestbest

)(best

)()(best

)(best

)(

)(best

)()(best

)()(

)(best

)(

<

+←

−+−+←

=<

=<

==

=

10

9

7

4

3

2

1

2211ω

Page 10: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

10www.hm.pjwstk.edu.pl

Decomposition of synchronous PSO algorithm on GPU

Page 11: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

11www.hm.pjwstk.edu.pl

Estimation of human configuration

Page 12: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

12www.hm.pjwstk.edu.pl

Framework

Page 13: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

13www.hm.pjwstk.edu.pl

Fitness function

Page 14: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

14www.hm.pjwstk.edu.pl

Rendering

Page 15: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

15www.hm.pjwstk.edu.pl

Human tracking

Page 16: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

16www.hm.pjwstk.edu.pl

•Notebook with Intel Core 2 Duo 2GHz, 3GB RAM•Notebook with nVidia GeForce GT 130M (4 multiprocessors, 1.5GHz, 512MB RAM)•PC with nVidia GeForce GTX 280 (30 multiprocessors, 1.3GHz, 1GB RAM)

Experiments

Page 17: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

17www.hm.pjwstk.edu.pl

Decomposition of synchronous PSO algorithm on GPU

Page 18: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

18www.hm.pjwstk.edu.pl

Experimental resultsComputation time [s]

x15.20.223.34#500, 5 it

x16.30.416.68#1000, 5 it

x16.50.8113.38#2000, 5 it

x16.81.5926.74#4000, 5 it

x15.70.396.12#500, 10 it

x16.40.7512.28#1000, 10 it

x16.41.4924.51#2000, 10 it

x16.22.9448.89#4000, 10 it

SpeedupGTX 280CPU

Page 19: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

19www.hm.pjwstk.edu.pl

• GPU implementation has achieved a speed-up of more then fifteen times then our CPU-based implementation• Tracking full human body can be preformed at 5 frames per second• Tracking speed is expected to be accelerated in near future by new generations of graphic cards

Conclusions

Page 20: GPU-Accelerated Tracking of the Motion of 3D Articulated ...€¦ · 1 GPU – Accelerated Tracking of the Motion of 3D Articu lated Figure Tomasz Krzeszowski Bogdan Kwolek Konrad

20www.hm.pjwstk.edu.pl

System wraz z biblioteką modułów dla zaawansowanej analizy i interaktywnej syntezy ruchu postaci ludzkiejProjekt współfinansowany ze środków Europejskiego Funduszu Rozwoju Regionalnego w ramach Programu Operacyjnego Innowacyjna Gospodarka 2007-2013. Działanie 1.3,

Poddziałanie 1.3.1.

PJWSTK - Wydział Zamiejscowy Informatyki w Bytomiu

Aleja Legionów 2, 41-902 Bytomtel.: +48 32 387-16-62 +48 32 387-16-62

fax: +48 32 389-01-31

Kontakt merytoryczny:prof. dr hab. inż. Konrad Wojciechowski (Kierownik

projektu)[email protected]

Kontakt administracyjny:Aleksandra Szołtysik

[email protected]

2010-12-28Warszawa