Motion Binary Patterns for Action Recognition

∗

Florian Baumann

, Jie Lao

, Arne Ehlers

and Bodo Rosenhahn

Institut f

ur Informationsverarbeitung, Leibniz Universit

at Hannover, Appelstraße 9a, 30167 Hannover, Germany

Department of Electronic Science and Technology, USTC, Hefei, Anhui 230027, P.R. China

Keywords:

Human Action Recognition, Volume Local Binary Patterns, Random Forest, Machine Learning, IXMAS,

KTH, Weizman.

Abstract:

In this paper, we propose a novel feature type to recognize human actions from video data. By combining the

beneﬁt of Volume Local Binary Patterns and Optical Flow, a simple and efﬁcient descriptor is constructed.

Motion Binary Patterns (MBP) are computed in spatio-temporal domain while static object appearances as

well as motion information are gathered. Histograms are used to learn a Random Forest classiﬁer which

is applied to the task of human action recognition. The proposed framework is evaluated on the well-known,

publicly available KTH dataset, Weizman dataset and on the IXMAS dataset for multi-view action recognition.

The results demonstrate state-of-the-art accuracies in comparison to other methods.

1 INTRODUCTION

Human action recognition is a complex area of com-

puter vision since static object characteristics, mo-

tion and time information have to be taken into ac-

count. Furthermore, actions are divided into human

actions, human-human interactions, human-object in-

teractions and group activities (Aggarwal and Ryoo,

2011). Due to environment variations such as moving

backgrounds, different view points or occlusions the

detection and classiﬁcation of actions is even more

difﬁcult. Additionally, each actor has its own style of

performing an action, leading to many variations in

the subject’s movement and a large intra-class varia-

tion (Aggarwal and Ryoo, 2011; Poppe, 2010).

Contribution. In this work, we address the problem

of recognizing actions performed by a single person,

e.g. boxing, clapping, waving, walking, running, jog-

ging. We suggest a simple and efﬁcient novel feature

type, namely Motion Binary Pattern (MBP), which

combines static object appearances as well as motion

information in the spatio-temporal space, in one de-

scriptor. An MBP is computed from three frames fol-

lowed by a histogram computation, leading to an in-

variance against different video lengths. Finally, the

histogram is used to learn a Random Forest classiﬁer.

The proposed approach is evaluated on the single-

view KTH dataset (Schuldt et al., 2004) and Weizman

∗

This work has been partially funded by the ERC within

the starting grant Dynamic MinVIP.

(Blank et al., 2005; Gorelick et al., 2007) dataset as

well as on the IXMAS (Weinland et al., 2006; Wein-

land et al., 2010) dataset for multi-view action recog-

nition. Figures 1, 2 and 3 show some example images

of the used datasets.

Related Work. The ﬁrst Local Binary Pattern on

Three Orthogonal Planes (LBP-TOP) was developed

by Zhao et al. (Zhao and Pietikainen, 2007) to clas-

sify dynamic textures while Mattivi and Shao applied

LBP-TOP (Mattivi and Shao, 2009; Shao and Mat-

tivi, 2010) to the task of human action recognition.

The authors reached 88.19% accuracy on the KTH

dataset and by combining Extended Gradient LBP-

TOP with Principal Component Analysis in (Mattivi

and Shao, 2009) they reached 91.25%. Best results

(92.69%) were achieved by combining Extended Gra-

dient LBP-TOP with Dollar’s detection method (Shao

and Mattivi, 2010).

Liu et al. (Liu and Yuen, 2010) proposed a boosted

EigenActions framework which calculates a spatio-

temporal information saliency map (ISM) by estimat-

ing pixel density functions. It has only 81.5% accu-

racy for the KTH dataset but reaches accuracies up to

98.3% on the Weizman dataset.

Using only single descriptors, the best result is not

more than 90% in the above mentioned papers. Bet-

ter results are obtained by Yeffet et al. (Yeffet and

Wolf, 2009), the average accuracy is a little more than

90%. And most recently, Kihl et al. (Kihl et al., 2013)

reached 93.4% with a series of local polynomial ap-

385

Baumann F., Lao J., Ehlers A. and Rosenhahn B..

Motion Binary Patterns for Action Recognition.

DOI: 10.5220/0004816903850392

In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 385-392

ISBN: 978-989-758-018-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

proximation of Optical Flow (SoPAF).

Many approaches have drawbacks, for instance

the amount of feature lead to ambiguities, cannot deal

with different video lengths or have a large feature

space. To overcome these issues, we suggest a new

feature that combines the capabilities of describing

static object appearances as well as motion informa-

tion, in a single descriptor.

2 METHOD

In this Section, Volume Local Binary Patterns and

Motion Binary Patterns are described. VLBPs have

become famous for describing features in the spatio-

temporal domain and are derived from simple Local

Binary Patterns.

Our proposed MBP is computed in the X-Y-T

space too and additionally takes the temporal step size

into account.

Finally, Section 2.3 brieﬂy describes the well-

known machine learning approach Random Forest by

Leo Breiman (Breiman, 2001).

2.1 Volume Local Binary Pattern

A Local Binary Pattern (LBP) was ﬁrst described in

(Ojala et al., 1994) for texture classiﬁcation. The orig-

inal LBP is computed in a 3 ×3 cell by comparing ev-

ery gray value to the center one. If the neighbor values

are larger than the center one, an 1 is assigned to the

corresponding position, otherwise 0. By computing a

3 × 3 - LBP the codeword length is 8 bit. This code-

word is interpreted as a binary word and converted to

a decimal number. Finally, a histogram of all occur-

ring numbers is built.

Since LBP features describe static object appear-

ances they are not suited for action recognition where

motion and time information should be taken into

account. A Volume Local Binary Pattern (VLBP)

is introduced in (Zhao and Pietikainen, 2007). It

is computed in the spatial and temporal domain and

Figure 1: Example images of the single-view KTH dataset

(Schuldt et al., 2004). The dataset contains six actions per-

formed by 25 people under different conditions.

Figure 2: Example images of the multi-view IXMAS

dataset (Weinland et al., 2006; Weinland et al., 2010). The

dataset contains 12 actions. Each action is performed three

times by 12 people.

Figure 3: Example images of the single-view Weizman

dataset (Blank et al., 2005; Gorelick et al., 2007). The

dataset contains nine actions performed by nine people.

able to recognize dynamic textures. In (Zhao and

Pietikainen, 2007), the authors deﬁne a radius around

the center point within the space-time volume from

three continuous frames to get neighboring pixels

rather than using a 3 × 3 cell from one frame.

The computation of a VLBP is similar to the LBP: if

the gray value of neighboring voxels within the space-

time volume is larger than that of the voxel’s center,

the corresponding position is assigned to an 1, other-

wise 0. By computing a VLBP the codeword length

is 24 bit, leading to 2

= 16777216 different pat-

terns. Similar to the LBP, a histogram of all occurring

patterns is computed. Often, this huge feature pool

leads to several ambiguities. To overcome this prob-

lem (Fehr, 2007; Topi et al., 2000) introduced an uni-

form LBP and demonstrate that the overall amount of

LBPs can be reduced to a small subset. Experiments

on object detection show that 90% of all possible pat-

terns belong to this subset. For our application of ac-

tion recognition uLBPs were unsuitable. Presumably,

the ﬁnal feature pool contains not enough information

to ﬁnd discriminative patterns in the spatio-temporal

space.

Figure 4 illustrates how to compute a VLBP, the

ﬁnal result is 2039583. A VLBP is computed from

three continuous frames followed by a histogram

computation of all occurring patterns. The histogram

is directly used to learn a Random Forest classiﬁer.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

386

Figure 4: Procedure of computing a Volume Local Binary

Pattern.

2.2 Motion Binary Pattern

In this Section, the characteristics and the computa-

tion of our proposed Motion Binary Pattern

is de-

scribed. Assuming that motion can be detected by

the change of pixel intensity values, MBPs are com-

puted from three frames and measure the motion be-

tween them. Similar to the Optical Flow (Horn and

Schunck, 1981), an MBP describes characteristics of

motion. Figure 5 explains the computation of an MBP

in three frames. The frames are divided into cells and

for three cells at the same position, the corresponding

values within one cell are compared. If the gray value

within one cell of the ﬁrst frame is larger than that in

the second frame, an 1 is assigned, otherwise 0. By

using the same method, the third frame is compared

to the second frame. The resulting two patterns are

combined by using an exclusive OR (XOR), leading

to the ﬁnal motion pattern.

Regarding the computation of all MBPs in three

frames, the number N of sampled patterns C

(x,y),

1 ≤ n ≤ N depends on the frame size, on the pat-

terns’s size and it’s step size. Each pattern represents

the motion between these frames while the binary val-

ues, especially the number of ones

, denoted

by their entry-wise 1-norm, can be interpreted as the

strength of motion. To distinguish between weak or

Source code available at http://www.tnt.uni-

hannover.de/staff/baumann/

Figure 5: Procedure of computing our proposed MBP in

three frames.

strong motions, a motion vector

I = (i

,...,i

)

is in-

troduced.

I is derived by a strength condition:

(

≥ S,

0, otherwise.

(1)

Each element of the motion vector is set to 1 if the

number of ones in the pattern C

(x,y) is higher than

or equal to the motion strength threshold S ∈ {1 .. .7}.

Thus,

I contains positions with a certain proportion of

motion. It describes the spatial-dependent changing

of intensity values from three frames and ideally only

positions with large changes in the pixel intensity are

assigned to 1.

An MBP descriptor is created by computing a his-

togram from all motion vectors of a video. Thus, the

histogram is directly used to construct a feature for

learning a Random Forest classiﬁer.

Temporal Variations. In order to learn features from

fast and slow motions, MBPs are not only computed

from three continuous frames. Obviously only fast ac-

tions could be recognized by deriving features from

continuous frames. In addition, four spatial scale

steps are deﬁned and MBPs are computed by incor-

porating these steps for shifting the pattern through

the space-time volume. A time step of t

= 1,2, 3,4

was empirically chosen. For the case of t

= 1, an

MBP is computed from three continuous frames. Ev-

ery second frame is collected for t

= 2. Respectively,

for t

= 3, 4 every third or fourth frame was chosen.

MotionBinaryPatternsforActionRecognition

387

Instead of creating a single histogram that can de-

scribe fast motions, four histograms are created to

characterize different kind of motions. These his-

tograms are concatenated and used to learn a Random

Forest classiﬁer.

2.3 Random Forest

In this part, a brief explanation of the theory behind

Random Forest is given. Random Forests were de-

veloped by Leo Breiman (Breiman, 2001) and com-

bine the idea of bagging (Breiman, 1996) with a ran-

dom feature selection proposed by Ho (Ho, 1995; Ho,

1998) and Amit (Amit and Geman, 1997). A Random

Forest consists of a collection of CART-like decision

trees h

, 1 ≤ t ≤ T :

{h(~x,Θ

)

t=1,...T

}

where {Θ

} is a bootstrap sample from the training

data. Each tree casts a vote on a class for the input

~x. The class probabilities are estimated by majority

voting and used to calculate the sample’s label y(~x)

with respect to a given feature vector~x:

y(~x) = argmax

∑

t=1

(~x)=c

(2)

The decision function h

(~x) returns the result class c

of one tree with the indicator function F:

(~x)=c

(

1, h

(~x) = c,

0, otherwise.

(3)

Random Forest has a high classiﬁcation accuracy and

can deal with large data sets for multiple classes with

outstanding time efﬁciency.

2.3.1 Classiﬁcation

Input descriptors are classiﬁed by passing them down

each tree until a leaf node is reached. The result class

is deﬁned by each leaf node and the ﬁnal decision is

determined by taking the class having the most votes

(majority vote), see Equation (2).

3 EXPERIMENTAL RESULTS

KTH: VLBPs and MBPs are evaluated on the well-

known and publicly available KTH dataset (Schuldt

et al., 2004) consisting of six classes of actions. Each

action is performed by 25 persons in four different

scenarios. The KTH dataset consists of 599 videos.

Similar to (O’Hara and Draper, 2012), a ﬁxed position

bounding box with a temporal window of 24 frames

is selected, based on annotations by Lui (Lui et al.,

2010). Presumably, a smaller number of frames is suf-

ﬁcient (Schindler and Van Gool, 2008). Furthermore,

the original training/testing splits from (Schuldt et al.,

2004) are used.

Weizman: In a second experiment we evaluate

MBPs on the well-established Weizman action dataset

(Blank et al., 2005; Gorelick et al., 2007). In our opin-

ion, the Weizman dataset is already solved since many

researchers report accuracies of 100%. However, in

recent publications (Li et al., 2013; Tian et al., 2013;

Yu et al., 2013) this dataset is still used to evaluate the

corresponding methods. In order to allow a compar-

ison to recent works and to show the beneﬁt of our

proposed method we evaluate Motion Binary Patterns

on this dataset too. The Weizman dataset consists of

nine actions while each action is performed by nine

different persons. We manually labeled the dataset

and used the bounding boxes for the classiﬁcation.

The bounding boxes are available for download at our

homepage

IXMAS: Additionally, we evaluated MBPs on the

IXMAS dataset for multi-view action recognition

(Weinland et al., 2006; Weinland et al., 2010). The

IXMAS dataset contains 12 classes of actions. Each

action is performed three times by 12 persons while

the body position and orientation is freely chosen

by the actor. The IXMAS dataset consists of 1800

videos. A 5-fold cross validation is used to get the

results.

3.1 Evaluation for Volume Local Binary

Patterns

Several strategies for computing VLBP values were

tested. Two different neighborhoods (eight and four

values) were compared, the inﬂuence of different

histogram ranges as well as the difference between

frame-by-frame learning and multi-frame learning

has been evaluated. Best results were achieved by

computing a 4-value VLBP with multi-frame learn-

ing (one histogram for all frames of a video is cre-

ated) and a histogram range of 400 bins. Figure 6(a)

shows the confusion matrix with an average accuracy

of 89.81% for the KTH dataset.

3.2 Evaluation for Motion Binary

Patterns

For computing an MBP, the motion strength threshold

has to be adjusted. This parameter strongly inﬂuences

the performance of the MBP. Furthermore, MBPs are

http://www.tnt.uni-hannover.de/staff/baumann/

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

388

(a) (b)

Figure 6: Confusion matrices for the KTH dataset, (a): ref-

erence method (VLBP) with 89.81% accuracy, (b) proposed

method (MBP) with 91.83% accuracy. Most confusions oc-

cur in similar actions like walking, running, jogging and

boxing, waving, clapping.

Figure 7: Confusion matrix for applying Motion Binary

Patterns to the IXMAS dataset. The average accuracy is

80.55%. Most confusions occur in similar actions like wav-

ing, punching and sitting down, getting up.

more sensitive to different image sizes. In this Sec-

tion, we tested the inﬂuence of these parameters and

compare the results to several recent approaches.

3.2.1 Inﬂuence of the Motion Strength

Threshold

Section 2.2 gives a brief explanation about the motion

strength threshold S ∈ {1 . .. 7}. Table 1 shows the

recognition accuracy when the threshold varies for a

frame size of 75 × 150. When the threshold is larger

than seven, there will be fewer non-zero values in the

MBP histogram. As listed in Table 1, recognizing ac-

curacy increases when the threshold becomes larger.

The highest accuracy 92.13% is achieved by using a

threshold of six, leading to the assumption that this

value is perfectly suited for the task of human action

recognition.

Figure 8: Confusion matrix for applying Motion Binary Pat-

terns to the Weizman dataset. The accuracy is 100.00%.

Table 1: Average accuracy for MBP on the KTH dataset

with different motion thresholds. A threshold of 6 is leading

to the best accuracy.

Threshold Average accuracy (%)

1 74.53

2 83.33

3 85.65

4 90.74

5 88.89

6 92.13

7 90.27

3.2.2 Inﬂuence of Different Frame Sizes

We report results on two approaches of changing

frame sizes, using a motion strength threshold of six.

In the ﬁrst experiment, all frames are resized to a

squared size. Table 2 shows that the average accuracy

tends to increase too when increasing the side length

of the square from 50 to 135 pixel, inspite of some

ﬂuctuations. However, when we raise the frame size

from 135 × 135 to 150 × 150, the accuracy does not

increase considerably.

In a second experiment all frames are resized to a rect-

angular size. Table 3 shows the accuracies for a pro-

portionally changed width and height. By this means,

we can achieve an accuracy of 92.13% when width

and height of the rectangles are 75 ×150. Figure 6(b)

presents the confusion matrix of this result. The ac-

curacy for hand-waving and hand-clapping is higher

than that by VLBP and the result for boxing can reach

100%. Unlike the VLBP, MBP does not do well in

recognizing running. MBP is less sensitive to fast

motions than VLBP and fast actions may puzzle the

descriptor. But MBPs are sensitive to weak motions.

From all our test results, wrong classiﬁcations mainly

MotionBinaryPatternsforActionRecognition

389

Table 2: Average accuracy for MBP on the KTH dataset by

squared frame sizes. The best accuracy was achieved by

taking a frame size of 135 × 135.

Size Average accuracy (%)

50 × 50 85.64

60 × 60 83.33

75 × 75 87.50

80 × 80 89.35

100 × 100 87.96

120 × 120 88.89

135 × 135 90.74

Table 3: Average accuracy for MBP on the KTH dataset by

different frame sizes. The best accuracy was achieved by

taking a frame size of 75 × 150.

Size Average accuracy(%)

64 × 128 89.35

128 × 64 87.96

75 × 150 92.13

150 × 75 87.50

80 × 160 89.35

happen on similar actions like walking, running, jog-

ging and boxing, waving, clapping, as showed in Fig-

ure 6(b).

3.2.3 Comparison to State-of-the-Art Methods

In this Section, we compare the proposed method to

several state-of-the-art works and show that the MBPs

are a very efﬁcient descriptor for action recognition.

KTH: The MBP achieves an accuracy of 92.13% on

the KTH dataset. Table 4 reports the accuracies of

our proposed MBP in comparison to other methods.

Motion Binary Patterns reach the highest accuracy

for original training-/testing split and is only slightly

lower than the best result with cross-validation.

IXMAS: Table 5 presents the results of MBP applied

to the IXMAS dataset for multi-view action recogni-

tion in comparison to other state-of-the-art methods.

Figure 7 shows the confusion matrix. The matrix also

reveals that most confusions occur at similar actions

like waving, punching and sitting down, getting up.

We used a motion threshold of six and window size

of 75 × 150. For this experiment we compare MBPs

to single- and multi feature approaches. Our result of

80.55%, see Table 5 is based on a 5-fold cross valida-

tion and is only slightly lower than the best result (Li

and Zickler, 2012).

Weizman: Figure 8 shows the confusion matrix for

applying MBPs to the Weizman dataset. The accuracy

Table 4: Comparison to recent approaches on the KTH

dataset with a single descriptor.

Name Accuracy (%)

(Kihl et al., 2013) 93.4

(Kihl et al., 2013) 91.5

(Yeffet and Wolf, 2009) 90.1

(Laptev et al., 2008) 91.8

Proposed method 92.1

(Schindler and Van Gool, 2008) 92.7

Table 5: Average accuracy for MBPs on the IXMAS dataset

in comparison to single- and multi-feature methods.

Name Accuracy (%)

(Wang et al., 2012) 76.50

(Wu et al., 2011) 78.02

Proposed method 80.55

(Li and Zickler, 2012) 81.22

is 100.00%. Table 6 presents a comparison to single-

and multi-feature methods. Several approaches report

perfect recognition accuracies.

4 CONCLUSIONS AND FUTURE

WORK

In this paper a novel feature type, namely Motion Bi-

nary Patterns (MBP) are proposed. MBPs combine

the advantages of Volume Local Binary Patterns to

gather static object information and Optical Flow to

obtain motion information. An MBP is computed

from three frames with a temporal shifted sliding win-

dow. The resulting histograms are used to learn a

Random Forest classiﬁer.

The proposed feature is evaluated on the well-known,

publicly available KTH dataset, Weizman dataset

and on the IXMAS multi-view dataset. The results

demonstrate state-of-the-art accuracies in comparison

to Volume Local Binary Patterns and to other single-

and multi feature methods. The source code for an

MBP computation is available at http://www.tnt.uni-

hannover.de/staff/baumann/.

Future Work Our plans for future work are to evalu-

ate Motion Binary Patterns on more complex datasets

like Hollywood (Laptev et al., 2008), Hollywood2

(Marszałek et al., 2009) or YouTube action dataset

(Liu et al., 2009). Furthermore, we plan to elimi-

nate the manual adjustment of the motion threshold

by introducing an entropy function that chooses pat-

terns with more discriminative power. Alternatively,

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

390

Table 6: Average accuracy for MBPs on the Weizman

dataset in comparison to single- and multi-feature methods.

Name Accuracy (%)

(Jhuang et al., 2007) 98.80

(Lin et al., 2009) 100.00

(Blank et al., 2005) 100.00

(Gorelick et al., 2007) 100.00

Proposed method 100.00

(Schindler and Van Gool, 2008) 100.00

we suggest to encode more information into the pat-

tern. For instance, all temporal shifted patterns could

be integrated into one ﬁnal histogram. Additionally,

more research is needed to choose the optimal cell

size of a Motion Binary Pattern. In this paper we sug-

gest to compute a MBP in a 3 × 3 cell but the results

might be improved by taking other cell sizes like 5×5

or 7 × 7 .

REFERENCES

Aggarwal, J. and Ryoo, M. (2011). Human activity analy-

sis: A review. ACM Computing Surveys, 43(3):16:1–

16:43.

Amit, Y. and Geman, D. (1997). Shape quantization and

recognition with randomized trees. Neural computa-

tion, 9(7):1545–1588.

Blank, M., Gorelick, L., Shechtman, E., Irani, M., and

Basri, R. (2005). Actions as space-time shapes. In

Computer Vision (ICCV), 10th International Confer-

ence on, pages 1395–1402.

Breiman, L. (1996). Bagging predictors. In Machine Learn-

ing, volume 24, pages 123–140.

Breiman, L. (2001). Random forests. Machine learning,

45(1):5–32.

Fehr, J. (2007). Rotational invariant uniform local binary

patterns for full 3d volume texture analysis. In Finnish

signal processing symposium (FINSIG).

Gorelick, L., Blank, M., Shechtman, E., Irani, M., and

Basri, R. (2007). Actions as space-time shapes. Pat-

tern Analysis and Machine Intelligence (PAMI), IEEE

Transactions on, 29(12):2247–2253.

Ho, T. K. (1995). Random decision forests. In Document

Analysis and Recognition, 1995., Proceedings of the

Third International Conference on, volume 1, pages

278–282. IEEE.

Ho, T. K. (1998). The random subspace method for con-

structing decision forests. Pattern Analysis and Ma-

chine Intelligence, IEEE Transactions on, 20(8):832–

844.

Horn, B. K. and Schunck, B. G. (1981). Determining optical

ﬂow. Artiﬁcial Intelligence, 17.

Jhuang, H., Serre, T., Wolf, L., and Poggio, T. (2007). A

biologically inspired system for action recognition. In

Computer Vision (ICCV), 11th International Confer-

ence on, pages 1–8. IEEE.

Kihl, O., Picard, D., Gosselin, P.-H., et al. (2013). Local

polynomial space-time descriptors for actions classi-

ﬁcation. In International Conference on Machine Vi-

sion Applications.

Laptev, I., Marszalek, M., Schmid, C., and Rozenfeld,

B. (2008). Learning realistic human actions from

movies. In Computer Vision and Pattern Recognition,

(CVPR). IEEE Conference on.

Li, R. and Zickler, T. (2012). Discriminative virtual views

for cross-view action recognition. In Computer Vision

and Pattern Recognition, (CVPR). IEEE Conference

on.

Li, W., Yu, Q., Sawhney, H., and Vasconcelos, N. (2013).

Recognizing activities via bag of words for attribute

dynamics. In Computer Vision and Pattern Recogni-

tion, (CVPR). IEEE Conference on, pages 2587–2594.

Lin, Z., Jiang, Z., and Davis, L. S. (2009). Recognizing

actions by shape-motion prototype trees. In Com-

puter Vision (ICCV), 12th International Conference

on, pages 444–451. IEEE.

Liu, C. and Yuen, P. C. (2010). Human action recognition

using boosted eigenactions. Image and vision comput-

ing, 28(5):825–835.

Liu, J., Luo, J., and Shah, M. (2009). Recognizing realis-

tic actions from videos “in the wild”. In Computer

Vision and Pattern Recognition, 2009. CVPR 2009.

IEEE Conference on, pages 1996–2003. IEEE.

Lui, Y. M., Beveridge, J., and Kirby, M. (2010). Action clas-

siﬁcation on product manifolds. In Computer Vision

and Pattern Recognition, (CVPR). IEEE Conference

on.

Marszałek, M., Laptev, I., and Schmid, C. (2009). Actions

in context. In Computer Vision and Pattern Recogni-

tion, (CVPR). IEEE Conference on.

Mattivi, R. and Shao, L. (2009). Human action recogni-

tion using lbp-top as sparse spatio-temporal feature

descriptor. In Computer Analysis of Images and Pat-

terns (CAIP).

O’Hara, S. and Draper, B. (2012). Scalable action recogni-

tion with a subspace forest. In Computer Vision and

Pattern Recognition, (CVPR). IEEE Conference on.

Ojala, T., Pietikainen, M., and Harwood, D. (1994). Perfor-

mance evaluation of texture measures with classiﬁca-

tion based on kullback discrimination of distributions.

In Pattern Recognition. Proceedings of the 12th IAPR

International Conference on.

Poppe, R. (2010). A survey on vision-based human action

recognition. Image and Vision Computing, 28(6):976

– 990.

Schindler, K. and Van Gool, L. (2008). Action snippets:

How many frames does human action recognition re-

quire? In Computer Vision and Pattern Recognition,

(CVPR). IEEE Conference on.

Schuldt, C., Laptev, I., and Caputo, B. (2004). Recogniz-

ing human actions: a local svm approach. In Pattern

Recognition. (ICPR). Proceedings of the 17th Interna-

tional Conference on.

MotionBinaryPatternsforActionRecognition

391

Shao, L. and Mattivi, R. (2010). Feature detector and de-

scriptor evaluation in human action recognition. In

Proceedings of the ACM International Conference on

Image and Video Retrieval.

Tian, Y., Sukthankar, R., and Shah, M. (2013). Spatiotem-

poral deformable part models for action detection. In

Computer Vision and Pattern Recognition, (CVPR).

IEEE Conference on.

Topi, M., Timo, O., Matti, P., and Maricor, S. (2000). Ro-

bust texture classiﬁcation by subsets of local binary

patterns. In Pattern Recognition. (ICPR). Proceedings

of the 15th International Conference on.

Wang, Z., Wang, J., Xiao, J., Lin, K.-H., and Huang, T.

(2012). Substructure and boundary modeling for con-

tinuous action recognition. In Computer Vision and

Pattern Recognition, (CVPR). IEEE Conference on.

Weinland, D.,

Ozuysal, M., and Fua, P. (2010). Making

action recognition robust to occlusions and viewpoint

changes,. In European Conference on Computer Vi-

sion (ECCV).

Weinland, D., Ronfard, R., and Boyer, E. (2006). Free

viewpoint action recognition using motion history vol-

umes. In Computer Vision and Image Understanding

(CVIU).

Wu, X., Xu, D., Duan, L., and Luo, J. (2011). Action recog-

nition using context and appearance distribution fea-

tures. In Computer Vision and Pattern Recognition,

(CVPR). IEEE Conference on.

Yeffet, L. and Wolf, L. (2009). Local trinary patterns for

human action recognition. In Computer Vision and

Pattern Recognition, (CVPR). IEEE Conference on.

Yu, T.-H., Kim, T.-K., and Cipolla, R. (2013). Uncon-

strained monocular 3d human pose estimation by ac-

tion detection and cross-modality regression forest. In

Computer Vision and Pattern Recognition, (CVPR).

IEEE Conference on, pages 3642–3649.

Zhao, G. and Pietikainen, M. (2007). Dynamic texture

recognition using local binary patterns with an ap-

plication to facial expressions. Pattern Analysis and

Machine Intelligence (PAMI), IEEE Transactions on,

29(6):915–928.

ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods

392