Dotacje na innowacje

Transcription

Dotacje na innowacje
Dotacje na innowacje
Model-Based 3D Human Motion Capture
Using Global-Local Particle Swarm Optimization
Tomasz Krzeszowski, Bogdan Kwolek, Konrad Wojciechowski
Polish-Japanese Institute of Information Technology
Human motion tracking - our approach
Human motion tracking problem
Objective:
Estimation of the current configuration of the human
body
Problems:
High dimensional search space
High variability in human appearance
Occlusions of the individual body parts
Approach to the problem:
Motion Capture
Magnetic and inertial sensors
Methods based on image analysis
Mutlicamera system (4 cameras)
GLPSO algorithm
3D Model of human body
Image analysis (background
detection, edges extraction, skin
detection)
Ground truth - Motion Capture
System
image comparison
.
.
.
image of human
silhouette
images with hypothesis
of human silhouette
PSO
Stochastic optimization technique
Each of particles represents a
hypothetical solution
Asynchronous PSO with global
neighborhood
xi,j
vi,j
yj
GLPSO
t-1
vi,j (t+1)
a)
yi,j
t
torso, head
arms
legs
torso, head
arms
legs
b)
torso, head
arms
legs
c)
torso, head
arms
legs
d)
Fitness function
Person extraction
Person segmentation
e)
torso, head
arms
estimated full
body pose
estimate of full
body pose
legs
estimate of
legs pose
legs
estimated full
body pose
Tracking errors
Errors of full body tracking
Sequence 1
w - weight, f - fitness function
(s- skin, p - segmentation, e - edges)
Human motion tracking
Sequence 2
#particles
#it.
100
10
average
error [mm]
79.41
300
10
74.54
46.24
77.84
50.31
100
20
78.46
50.10
80.33
47.99
300
20
70.28
41.69
73.86
45.58
100
10
71.19
33.88
73.88
35.58
300
10
64.14
25.74
64.52
23.36
100
20
68.03
28.31
67.01
27.72
300
20
62.46
23.77
62.59
22.45
PSO
GLPSO
std dev.
[mm]
50.19
average
error [mm]
82.25
std dev.
[mm]
53.48
Errors of right leg tracking
PSO
59,88
average
error [mm]
50,13
63,01
56,87
44,84
27,17
20
95,57
84,72
70,4
44,12
20
93,25
77,91
72,39
36,44
#particles
#it.
100
20
average
error [mm]
68,08
300
20
100
300
right knee
right tibia
GLPSO
std dev. [mm]
std dev. [mm]
Conclusions
Effective algorithm for human motion tracking
Average tracking error of 62 mm for the full body
7th Int. Conf. on Computer Recognition Systems CORES 2011
Projekt współfinansowany ze śrdoków Europejskiego Funduszu Rozwoju Regionalnego w ramach Programu
Operacyjnego Innowacyjna gospodarka 2007-2013. Działanie 1.3, Poddziałanie 1.3.1.
33,18

Similar documents