Amsterdam, The Netherlands July 06-10, 2013 Real World Applications: RWA4. Room: 02A00 10:40 – 12:20 Session Chair: Alexandros Andre Chaaraoui (University of Alicante, Spain)
Jan 15, 2015
Amsterdam, The Netherlands July 06-10, 2013
Real World Applications: RWA4.
Room: 02A00 10:40 – 12:20
Session Chair: Alexandros Andre Chaaraoui (University of Alicante, Spain)
ALEXANDROS ANDRE CHAARAOUI AND FRANCISCO FLÓREZ-REVUELTA
HUMAN ACTION RECOGNITION
OPTIMIZATION BASED ON EVOLUTIONARY FEATURE
SUBSET SELECTION
… …
Amsterdam, July 6-10, 2013
Gen
etic
an
d
Evolu
tion
ary
C
om
pu
tatio
n
Con
fere
nce 2
01
3
3
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
Contents1. Introduction2. Radial Summary Feature3. Evolutionary Feature Subset
Selection4. Human Action Recognition
Method5. Experimentation & Results6. Conclusions7. ReferencesQ & A and Discussion
4
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
1. Introduction
Motivation and starting point Recognition of actions such as walking,
jumping or falling. Requirements:
High and stable recognition ratesReal-time suitability
Proposal of a visual feature with reduced extraction cost and low dimensionality
Feature subset selection
5
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
2. Radial Summary Feature
Human Silhouettes Relatively simple extraction
process Rich shape information Contour points
Radial Summary feature proposal Spatial alignment Feature
selection Low dimensionality, reduced
extraction cost, … Fig 1: Sample silhouette of the MuHAVi dataset [1].
6
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
2. Radial Summary Feature
Fig 2: Overview of the proposed Radial Summary feature.
7
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
3. Evolutionary Feature Subset Selection
Binary selection using a genetic algorithm Binary individual representation:
Active radial bin: uj = 1
Disabled radial bin: uj = 0
Random initial population (but one with all selected)
Fitness based on the evaluation of the feature Individuals with less active bins are favoured One-point crossover combination operator with
ranking selection Flip bit mutation operator Convergence termination criteria
8
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
4. Human Action Recognition Method
Pose Representations
Bag-of-Key-Poses Model
Sequences of Key Poses
9
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
4. Human Action Recognition Method
Learning based on Bag-of-Key-Poses Model The available pose representations
are reduced to a representative subset of key poses
We use the K-means clustering algorithm
10
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
4. Human Action Recognition Method
Sequence recognition Sequences of key poses Nearest-neighbour key poses Sequence matching (dynamic time
warping)
Fig 3: Sequences of key poses.
11
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
5. Experimentation & Results
Tested on the MuHAVi-MAS Dataset [1]
Two versions with 14 and 8 actions Manually Annotated Silhouettes Leave-one-actor-out (LOAO) and leave-one-
sequence out (LOSO) cross validations
Dataset Test Chaaraoui et al.
[2]
Radial Summar
y
Feature Selectio
n
State of the Art Rate [3]
MuHAVi-14
LOSO 94.1% 95.6% 98.5% 91.9%
MuHAVi-14
LOAO 86.8% 91.2% 94.1% 77.9%
MuHAVi-8 LOSO 98.5% 100% 100% 98.5%
MuHAVi-8 LOAO 95.6% 97.1% 100% 85.3%
12
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
5. Experimentation & Results Result of the feature
selection ~47% feature size
reduction
~14% temporal reduction
96 FPS overall recognition rate Fig 4: Resulting feature subset
selection of the MuHAVi-14 LOSO cross validation test (dismissed radial bins are shaded in gray).
13
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
6. Conclusions
Conclusions An evolutionary algorithm has been applied to
optimize action recognition. An appropriate feature for feature subset
selection has been proposed. We demonstrated that a guided selection of
feature elements can improve the recognition rate and reduce the computational cost.
Future work Real-valued weights instead of binary selection Action-class specific feature selection
14
© A
lexandro
s A
ndre
Ch
aara
oui and F
ranci
sco F
lóre
z-R
evu
elt
a
(GEC
CO
’13
)
7. References
[1] Singh, S., Velastin, S.A., Ragheb, H.: Muhavi: A multicamera human action video dataset for the evaluation of action recognition methods. In: 2010 Seventh IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), pp. 48–55 (2010)
[2] Chaaraoui, A.A., Climent-Perez, P., Florez-Revuelta, F.: An Efficient Approach for Multi-view Human Action Recognition based on Bag-of-Key-Poses. In Salah, A., ed.: Human Behavior Understanding. Lecture Notes in Computer Science. Springer Berlin / Heidelberg (2012)
[3] A. Eweiwi, S. Cheema, C. Thurau, and C. Bauckhage. Temporal key poses for human action recognition. In Computer Vision Workshops (ICCV Workshops), IEEE International Conference on, pp. 1310-1317 (2011)
15 Q & A and Discussion
ALEXANDROS ANDRE CHAARAOUI AND FRANCISCO FLÓREZ-REVUELTA
HUMAN ACTION RECOGNITION
OPTIMIZATION BASED ON EVOLUTIONARY FEATURE
SUBSET SELECTION
… …
Amsterdam, July 6-10, 2013
Gen
etic
an
d
Evolu
tion
ary
C
om
pu
tatio
n
Con
fere
nce 2
01
3