Anticipating Suspicious Actions using a Small Dataset of Action

Templates

Renato Baptista, Michel Antunes, Djamila Aouada and Bj

orn Ottersten

Interdisciplinary Centre for Security, Reliability and Trust (SnT), University of Luxembourg, Luxembourg, Luxembourg

Keywords:

Action Templates, Alarm Generation, Skeleton.

Abstract:

In this paper, we propose to detect an action as soon as possible and ideally before it is fully completed. The

objective is to support the monitoring of surveillance videos for preventing criminal or terrorist attacks. For

such a scenario, it is of importance to have not only high detection and recognition rates but also low time

latency for the detection. Our solution consists in an adaptive sliding window approach in an online manner,

which efﬁciently rejects irrelevant data. Furthermore, we exploit both spatial and temporal information by

constructing feature vectors based on temporal blocks. For an added efﬁciency, only partial template actions

are considered for the detection. The relationship between the template size and latency is experimentally

evaluated. We show promising preliminary experimental results using Motion Capture data with a skeleton

representation of the human body.

1 INTRODUCTION

Many surveillance systems are composed of cameras

acquiring videos from speciﬁc locations and monito-

red by people (e.g. a security team) for detecting sus-

picious events and actions. It is a challenging task

to manually monitor video feeds 24/7. As a conse-

quence, only after criminal or terrorist attacks occur,

the recorded surveillance data is actually used for ana-

lyzing what happened at that speciﬁc moment.

Nowadays, there are many visual surveillance sy-

stems that apply computer vision techniques for auto-

matically detecting suspicious occurrences, including

“human violence” recognition and detection (Bilin-

ski and Bremond, 2016; Datta et al., 2002). General

action recognition and detection is a largely investiga-

ted topic by the computer vision community, showing

very promising results (Du et al., 2015; Gkioxari and

Malik, 2015; Wang and Schmid, 2013; Wang et al.,

2015; Papadopoulos et al., 2017). However, a major

concern in security applications is not only the accu-

rate detection and recognition of particular events or

actions, but also the time latency required for achie-

ving it. Many of the existing works are designed

for action recognition (Bilen et al., 2016; Du et al.,

2015; Fernando et al., 2016) and for ofﬂine action

detection (Gaidon et al., 2011; Gkioxari and Malik,

2015; Jain et al., 2014; Tian et al., 2013; Wang et al.,

2015; Papadopoulos et al., 2017). These methods re-

... ... ...

Figure 1: Alarm generation for a suspicious action during

an online stream from a video surveillance camera. The

green color represents a normal action and the blue color

the ground truth information. The goal is to detect the sus-

picious action as soon as possible and ideally before it is

fully completed. The orange ﬂags represent the alarms that

were generated during the video using our approach, and

the red ﬂag corresponds to the ﬁrst alarm that was genera-

ted for the suspicious action. Images used were extracted

from the CAVIAR dataset

quire that the action to be recognized is completely

acquired before the detection can be accomplished.

However, even if the recognition accuracy is 100%,

it is not recommended to use these approaches in se-

curity and surveillance applications, because an alarm

can only be issued after the event has occurred. An

alternative to these approaches would be an online

action detection approach, where the objective is to

detect an action as soon as it happens and ideally be-

fore the action is fully completed. Enabling to detect

an action with low latency can be useful in many vi-

http://homepages.inf.ed.ac.uk/rbf/CAVIAR

380

Baptista, R., Antunes, M., Aouada, D. and Ottersten, B.

Anticipating Suspicious Actions using a Small Dataset of Action Templates.

DOI: 10.5220/0006648703800386

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 5: VISAPP, pages

380-386

ISBN: 978-989-758-290-5

deo surveillance applications.

Recently, Hoai and De la Torre (Hoai and De la

Torre, 2012; Hoai and De la Torre, 2014) proposed

a learning formulation based on a structured output

Support Vector Machine (SVM) to recognize partial

events, resulting in an early detection. Detecting an

action in an online manner is not a trivial problem due

to the unpredictability of real word scenarios (Geest

et al., 2016; Li et al., 2016). Geest et al. (Geest

et al., 2016) proposed a more realistic dataset for on-

line action detection. The dataset consists of real life

actions that were professionally recorded from six re-

cent TV series (Geest et al., 2016). Li et al. (Li et al.,

2016) presented a deep learning based architecture

that allows to detect and recognize actions in an on-

line manner. The authors show very promising results

on a large dataset containing video streams. The ma-

jor weakness of this approach is that, being based on

deep learning methods, it requires a large amount of

data for training the overall architecture. Large data-

sets of criminal and terrorist attacks are usually not

available, because they only occur sporadically. This

means that training a deep architecture with many lay-

ers and parameters is highly challenging.

In this work, the objective is to detect suspicious

events and actions as soon as possible and ideally be-

fore they are fully completed. Note that, the goal of

this approach is to support the video surveillance rat-

her than being a completely automated system. We

propose to use an adaptive sliding window, which efﬁ-

ciently rejects irrelevant data during streaming, toget-

her with a histogram based video descriptor and a ne-

arest neighbor assignment, which have the advantage

of efﬁcient iterative computations. As the objective

is to support video surveillance and security monito-

ring applications, one prefers to have many alerts with

higher probabilities of detecting a suspicious action

with very low latency, Figure 1 illustrates an example

of a detection scenario using our approach. Similarly

to (Li et al., 2016; Meshry et al., 2015; Sharaf et al.,

2015), we use a skeleton representation of the human

body as it is robust to scale, rotation and illumination

changes, and it can be computed at high frame rate,

allowing real-time computations (Han et al., 2017).

Such a human body representation can be extracted

from human pose estimation algorithms (Pishchulin

et al., 2016). In contrast to (Li et al., 2016), our goal

is to avoid the requirement of a large amount of anno-

tated data. To that end, we propose to use a small da-

taset of suspicious actions and also to construct tem-

plate actions using different percentages of the action

to be detected. Furthermore, we present an analysis of

the inﬂuence of using partial information of the action

to be detected, showing that it is possible to achieve

competitive detection results when compared to using

the complete action sequence.

In summary, the contributions of this work are: 1)

an efﬁcient algorithm for detecting actions using a

small dataset of template actions with low latency;

and 2) an analysis of the time needed to detect an

action using partial information of the action tem-

plate.

The paper is organized as follows: in Section 2,

we provide a brief introduction of the skeleton repre-

sentation of the human body and the problem formu-

lation for the proposed approach. Section 3 proposes

the online action detection method and how to con-

struct action templates. In Section 4, we describe and

discuss the experimental results and Section 5 conclu-

des the paper.

2 BACKGROUND & PROBLEM

FORMULATION

In this section, we introduce the skeleton human body

representation that is used throughout the paper. Let

us assume that a human action video is represented by

the spatial positions of the body joints (Antunes et al.,

2016; Baptista et al., 2017a; Baptista et al., 2017b;

Vemulapalli et al., 2014). A skeleton S = [j

,··· , j

]

is deﬁned using N joints, and each joint is represented

by its 3D coordinates j = [ j

, j

]

, where j

x,y,z

∈ R

and T denotes the matrix transpose. A human ske-

leton sequence is represented by H = {S

,··· , S

where F is the total number of frames. In order to

normalize each skeleton in a way that the size of each

body part is in correspondence, a spatial registration

is done by transforming each skeleton S, such that the

world coordinate system is placed at the hip center

and rotated in such a manner that the projection of the

vector from the left hip to the right hip is parallel to

the x-axis (Vemulapalli et al., 2014). Figure 2 shows

an example of the normalized skeleton with respect to

the world coordinate system placed at the hip center

and the enumerated skeleton joints. Each skeleton S

is then represented by the 3D normalized coordinates

of N joints as a vector of size 3N. In this work, we

adopt the same approach as in (Chu et al., 2012), and

a sequence H is represented using a bag of temporal

words model (Sivic and Zisserman, 2003; Yuan et al.,

2011). In this model, the codebook is built by using k-

means in order to group similar feature vectors. Each

skeleton S from H is discretized into histograms ac-

cording to the k-entry dictionary. Then, the resulting

representation of the sequence H is deﬁned by the fea-

ture vector ϕ(H), which is the cumulative summation

of all the individual histograms of the sequence H.

Anticipating Suspicious Actions using a Small Dataset of Action Templates

381

Figure 2: Representation of the normalized skeleton with

respect to the world coordinate system placed at the hip cen-

ter (joint number 0). L and R stand for the left and right side

of the human body, respectively.

Considering an action template A

of length T ,

and a subsequence of an input human action H

⊂ H

starting from frame s to the current frame t, i.e., H

,··· , S

} such that 1 < s,t < F, the objective is to

estimate the starting point ˆs of the action of interest

such as (Chu et al., 2012):

ˆs = argmin

d(ϕ(H

),ϕ(A

)) s.t. t −s ≥ L, (1)

where L is the minimum length of the interval of in-

terest and d is a distance function measuring dissi-

milarity between histograms. In (Chu et al., 2012),

equation (1) is solved considering the full sequence H,

while in the problem at hand, the objective is to ﬁnd a

solution in an online manner. To that end, we propose

to follow an efﬁcient adaptive sliding window strategy

combined with: 1) adding temporal information to the

spatial feature representation ϕ(·); and 2) decreasing

the amount of data required from the action template.

The proposed approach is detailed in Section 3.

3 PROPOSED APPROACH

We herein describe the proposed method for online

action detection and how action templates are deﬁned.

First of all, the temporal information is added by

aggregating a consecutive number b of skeletons S to-

gether as a new (b ×3N)-dimensional vector, where b

deﬁnes the number of considered consecutive frames,

also known as the temporal block size. Then, the fe-

ature representation of the sequence H using b is re-

presented by ϕ

(H).

To solve (1), we compare the values of the dis-

tance function d for the two intervals w

= [s,t + 1],

0 150 300 450 600 750

Time (frames)

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Cost

min

Action of Interest

Figure 3: Representation of the minimal distance d

min

over

time. The blue and red lines represent the cost for each

temporal window w

and w

, respectively. The yellow line

represents the smallest distance between the distances cor-

responding to the two temporal windows w

and w

over

time. τ deﬁnes the threshold to validate a detection.

and w

= [s + 1,t + 1] corresponding to the two sub-

sequences H

t+1

and H

t+1

s+1

, respectively. We denote

by d

and d

the resulting distances with respect to

the template action A

. The distance function d is

the Euclidean distance between histograms. The mi-

nimum of the two, d

min

= min(d

) is saved until

a new minimum is found, and the start point is accor-

dingly updated. This means that, if d

min

is obtained

from w

, the start point will be increased by one for

the next time instant t + 2. In this case, the method

rejects irrelevant data while it searches for the best

start point of the action of interest. The minimal dis-

tance vector d

min

is then the stored values of d

min

over

time, i.e., d

min

= [d

min

]. As a temporal function, d

min

starts to decrease as soon as the action of interest

occurs, which means that in that interval the num-

ber of generated alarms increases signiﬁcantly. An

action is considered detected when d

min

decreases by

a number of τ consecutive blocks. For every time

that the threshold τ is met, an alarm is generated. Fi-

gure 3 illustrates the relation of the distance function

over time for the two temporal windows w

and w

and d

min

. As shown in Figure 3, the action is detected

as soon as d

min

decreases for a consecutive number

of blocks τ. Note that, in this example, d

min

is decre-

asing during the ground truth interval, which means

that while the action is happening, a relative number

of alarms are generated for each time that the thres-

hold τ is met. Then, the alarm of interest is the ﬁrst

alarm that is generated within the ground truth inter-

val.

As the objective is to detect an action as soon

as possible with low latency, using subsequences of

the template action can be advantageous to detect an

action before it is fully completed since less data is

used. Therefore, we propose to use partial action tem-

plates from the full action A

. We deﬁne subsequen-

ces A

⊂ A

, for 0 < p ≤ 1. Similarly to the ag-

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

382

gregated feature vectors ϕ

deﬁned for H, for each

subsequence A

, a histogram is created by accumu-

lating individual blocks, where each block of size b

is assigned according to the k-entry codebook. Then,

the representation of an action template ϕ

) is

the average of all histograms of the same suspicious

action.

4 EXPERIMENTAL RESULTS

In this section, we evaluate the detection and the time

to detection performance on the public CMU-Mocap

dataset

. From this dataset we select a set of acti-

ons from different subjects that fall into two different

groups, the normal and the suspicious actions. The

actions “walking”, “standing” and “looking around”

are considered as normal actions, and the actions

“running”, “punching” and “kicking” are considered

as suspicious actions. Each sequence from the dataset

was recorded at 120 frames per second (fps). To make

it more realistic for surveillance purposes, we downs-

ample each sequence by a factor of 4, resulting in a 30

fps sequence. Figure 4 shows an example of the ske-

leton representation of human body while a suspici-

ous action is happening and the corresponding alarm

generation.

0 100 200 300 400 500 600

Time (frames)

0.2

0.4

0.6

0.8

Cost

min

Figure 4: Skeleton representation of the human body while

a suspicious action is happening, in this case: “Punching”.

The orange arrows represent the alarms that were generated

over time and the red arrow is the alarm of interest which is

the ﬁrst alarm generated within the ground truth.

4.1 Blocks vs. Threshold

In order to evaluate the detection and the time needed

to detect an action, we simulate an input video stream

by randomly concatenating normal actions before and

after the suspicious action (action of interest). First,

http://mocap.cs.cmu.edu

we start by evaluating the relation between the block

size b and the detection accuracy. Moreover, an eva-

luation of the time needed to detection is also done

in order to see how much X % of the action was nee-

ded to complete a detection. We determine a correct

detection if the minimal distance function d

min

decre-

ases by τ consecutive blocks and if the resulting de-

tection is within the ground truth interval. A range of

different values for the parameters were tested and for

the best detection accuracy with the lowest latency.

We ﬁx b = 3 frames and τ = 2 for the following expe-

riments. Figure 5(a) and 5(d) illustrate the detection

accuracy using different lengths of the subsequence

of the template action. Figure 5(b) and 5(e) show

an evaluation of the time needed to detect an action

for the different subsequences of the template action,

where the lowest latency obtained is around 12%. In

these experiments, it is shown that the proposed ap-

proach can be applied on a wide range of applications

for video surveillance. For applications that require

fast detection, the value of the threshold τ should be

lower and on the other hand, if the application requi-

res a more precise detection and the latency is not a

constraint, the threshold value should be higher. For

example, in a multi screen surveillance monitoring

system, when detection occurs, an alarm is ﬂagged

for the speciﬁc screen where the action is happening.

This will get the attention of the security guard, to

then analyse the action and make a decision. Such

applications allow a less sensitive system, where a

false alarm can be generated without compromising

the security, since it only alerts the security guard to

look at the camera where the alarm was generated. Fi-

gure 5(c) and 5(f) show the average number of alarms

that were generated per video. Furthermore, it also

shows the number of positive alarms per video, where

each color represents the number of correct detecti-

ons per video for the different b and τ. Note that,

using b = 2 until b = 5, the latency is between 12%

and 22% and the detection accuracy is between 68%

and 81%. Considering this and depending on the ap-

plication, these parameters can be tuned with respect

to the desired number of generated alarms per video.

Increasing the number of temporal blocks will reduce

the number of alarms that are generated per video.

4.2 Impact of the Size of the Template

Action

In order to evaluate the impact of the size of the sub-

sequences of the template action, we use the follo-

wing percentages of the initial part of the action p =

and the full action. We propose these subse-

quence lengths in order to understand the relations-

Anticipating Suspicious Actions using a Small Dataset of Action Templates

383

2 3 4 5 7 9

Block Size b

0.5

0.6

0.7

0.8

0.9

Average Detection (%)

p = 1/10

p = 1/4

p = 1/2

p = 1

(a) Average detection for τ = 2 blocks.

2 3 4 5 7 9

Block Size b

0.1

0.2

0.3

0.4

0.5

Time To Detection (%)

p = 1/10

p = 1/4

p = 1/2

p = 1

(b) Time to detection for τ = 2 blocks. (c) Average of positive alarms that were

generated per video for τ = 2 blocks. The

blue dotted line represents the total num-

ber of generated alarms per video (posi-

tive and negative alarms).

2 3 4 5 6

0.5

0.6

0.7

0.8

0.9

Average Detection (%)

p = 1/10

p = 1/4

p = 1/2

p = 1

(d) Average detection for b = 3 frames.

2 3 4 5 6

0.1

0.2

0.3

0.4

0.5

0.6

Time To Detection (%)

p = 1/10

p = 1/4

p = 1/2

p = 1

(e) Time to detection for b = 3 frames. (f) Average of positive alarms that were

generated per video for b = 3 frames.

The blue dotted line represents the to-

tal number of generated alarms per video

(positive and negative alarms).

Figure 5: In the ﬁrst row of images we ﬁxed τ = 2 blocks and computed the average detection, the time to detection and the

average number of alarms and also the number of positive alarms that were generated per video for the different temporal

blocks b, respectively. In the second row we proceed with the same experiments, ﬁxing the block size b = 3 frames for

different τ blocks.

hip between the size of the template action and the

time needed to detect the action. Figure 6 shows

the average detection accuracy for the different sub-

sequences of the template action. The best detection

accuracy obtained is for b = 3 frames and for p =

This means that using half of the action as a subse-

quence of the template action, we achieve competitive

detection results with low latency. This setup can be

advantageous for applications where the action needs

to be detected with low latency. In addition, using

only a percentage of the initial part of an action can

decrease the time needed to detect an action, resulting

in a faster detection. Table 1 shows the detection

accuracy for each suspicious action separately. Note

that, for the action “Punching” the detection is hig-

her due to the fact that this action is more discrimina-

tive for the upper limbs when compared with the other

Table 1: Detection accuracy (%) for the following scena-

rio: b = 3 frames and τ = 2 blocks, for p =

and 1.

Actions

Running 61.23 44.95 70.82 79.87

Kicking 76.7 81.73 84.17 60.76

Punching 95.83 90.08 91.89 90.67

two suspicious actions. One possible way to increase

the detection accuracy for all actions would be a more

robust representation, such as the relative position of

the joints or a Fourier Temporal Pyramid (Vemula-

palli et al., 2014) representation. We did not imple-

ment these representations to avoid the complexity of

the descriptors in order to have a better understanding

of the performance and characteristics of the propo-

sed method.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

384

Percentages p of the action

Average Detection (%)

b = 2

b = 3

b = 4

b = 5

b = 7

b = 9

Figure 6: Average detection for the different subsequences

with length p of the template action for τ = 2.

5 CONCLUSIONS

In this paper, we proposed an online method to detect

suspicious actions with low latency. This method is

based on an adaptive sliding window which efﬁciently

rejects irrelevant data during streaming. We explored

the feature representation of a subsequence using the

spatial and temporal information of the video stream.

Furthermore, we evaluated the relationship between

the size of the template action and latency, where we

conclude that using half of the action as a template

action, the detection accuracy and the time needed to

detect the action achieve competitive and promising

results compared to using the full action as a tem-

plate. We also observed that tuning the parameters,

the method can be used for different setups of video

surveillance. Next, we intend to use real surveillance

videos coupled with a robust human pose detection

approach, e.g. (Pishchulin et al., 2016).

ACKNOWLEDGEMENTS

This work has been partially funded by the Natio-

nal Research Fund (FNR), Luxembourg, under the

CORE project C15/IS 10415355/3D-ACT/Bj

orn Ot-

tersten. This work was also supported by the Euro-

pean Union‘s Horizon 2020 research and innovation

project STARR under grant agreement No.689947.

REFERENCES

Antunes, M., Baptista, R., Demisse, G., Aouada, D., and

Ottersten, B. (2016). Visual and human-interpretable

feedback for assisting physical activity. In European

Conference on Computer Vision (ECCV) Workshop on

Assistive Computer Vision and Robotics Amsterdam,.

Baptista, R., Antunes, M., Aouada, D., and Ottersten, B.

(2017a). Video-based feedback for assisting physical

activity. In International Joint Conference on Compu-

ter Vision, Imaging and Computer Graphics Theory

and Applications (VISAPP).

Baptista, R., Antunes, M., Shabayek, A. E. R., Aouada, D.,

and Ottersten, B. (2017b). Flexible feedback system

for posture monitoring and correction. In IEEE In-

ternational Conference on Image Information Proces-

sing (ICIIP).

Bilen, H., Fernando, B., Gavves, E., Vedaldi, A., and Gould,

S. (2016). Dynamic image networks for action recog-

nition. In The IEEE Conference on Computer Vision

and Pattern Recognition (CVPR).

Bilinski, P. and Bremond, F. (2016). Human violence re-

cognition and detection in surveillance videos. In Ad-

vanced Video and Signal Based Surveillance (AVSS),

2016 13th IEEE International Conference on, pages

30–36. IEEE.

Chu, W.-S., Zhou, F., and De la Torre, F. (2012). Unsu-

pervised temporal commonality discovery. In Euro-

pean Conference on Computer Vision, pages 373–387.

Springer Berlin Heidelberg.

Datta, A., Shah, M., and Da Vitoria Lobo, N. (2002).

Person-on-person violence detection in video data. In

Proceedings of the 16 th International Conference on

Pattern Recognition (ICPR’02) Volume 1 - Volume

1, ICPR ’02, pages 10433–, Washington, DC, USA.

IEEE Computer Society.

Du, Y., Wang, W., and Wang, L. (2015). Hierarchical recur-

rent neural network for skeleton based action recog-

nition. In The IEEE Conference on Computer Vision

and Pattern Recognition (CVPR).

Fernando, B., Gavves, E., Oramas, J., Ghodrati, A., and

Tuytelaars, T. (2016). Rank pooling for action recog-

nition. IEEE Transactions on Pattern Analysis and

Machine Intelligence.

Gaidon, A., Harchaoui, Z., and Schmid, C. (2011). Actom

Sequence Models for Efﬁcient Action Detection. In

CVPR 2011 - IEEE Conference on Computer Vision

& Pattern Recognition, pages 3201–3208, Colorado

Springs, United States. IEEE.

Geest, R. D., Gavves, E., Ghodrati, A., Li, Z., Snoek, C.,

and Tuytelaars, T. (2016). Online action detection.

CoRR, abs/1604.06506.

Gkioxari, G. and Malik, J. (2015). Finding action tubes.

Han, F., Reily, B., Hoff, W., and Zhang, H. (2017). Space-

time representation of people based on 3d skeletal

data: A review. Computer Vision and Image Under-

standing, pages –.

Hoai, M. and De la Torre, F. (2012). Max-margin early

event detectors. In Proceedings of IEEE Conference

on Computer Vision and Pattern Recognition.

Hoai, M. and De la Torre, F. (2014). Max-margin early

event detectors. International Journal of Computer

Vision, 107(2):191–202.

Jain, M., van Gemert, J. C., J

egou, H., Bouthemy, P., and

Snoek, C. G. M. (2014). Action localization by tube-

Anticipating Suspicious Actions using a Small Dataset of Action Templates

385

lets from motion. In IEEE Conference on Computer

Vision and Pattern Recognition.

Li, Y., Lan, C., Xing, J., Zeng, W., Yuan, C., and Liu, J.

(2016). Online human action detection using joint

classiﬁcation-regression recurrent neural networks.

European Conference on Computer Vision.

Meshry, M., Hussein, M. E., and Torki, M. (2015). Action

detection from skeletal data using effecient linear se-

arch. CoRR, abs/1502.01228.

Papadopoulos, K., Antunes, M., Aouada, D., and Ottersten,

B. (2017). Enhanced trajectory-based action recogni-

tion using human pose. In IEEE International Confe-

rence on Image Processing (ICIP).

Pishchulin, L., Insafutdinov, E., Tang, S., Andres, B., An-

driluka, M., Gehler, P. V., and Schiele, B. (2016). Dee-

pcut: Joint subset partition and labeling for multi per-

son pose estimation. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 4929–4937.

Sharaf, A., Torki, M., Hussein, M. E., and El-Saban, M.

(2015). Real-time multi-scale action detection from

3d skeleton data. In Applications of Computer Vi-

sion (WACV), 2015 IEEE Winter Conference on, pa-

ges 998–1005. IEEE.

Sivic, J. and Zisserman, A. (2003). Video Google: A text

retrieval approach to object matching in videos. In

IEEE International Conference on Computer Vision,

volume 2, pages 1470–1477.

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

poral deformable part models for action detection. In

Proceedings of the 2013 IEEE Conference on Compu-

ter Vision and Pattern Recognition, CVPR ’13, pages

2642–2649, Washington, DC, USA. IEEE Computer

Society.

Vemulapalli, R., Arrate, F., and Chellappa, R. (2014). Hu-

man action recognition by representing 3d skeletons

as points in a lie group. In Proceedings of the IEEE

Conference on Computer Vision and Pattern Recogni-

tion.

Wang, H. and Schmid, C. (2013). Action recognition with

improved trajectories. In IEEE International Confe-

rence on Computer Vision, Sydney, Australia.

Wang, Z., Wang, L., Du, W., and Qiao, Y. (2015). Exploring

ﬁsher vector and deep networks for action spotting. In

The IEEE Conference on Computer Vision and Pattern

Recognition (CVPR) Workshops.

Yuan, J., Liu, Z., and Wu, Y. (2011). Discriminative video

pattern search for efﬁcient action detection. Pattern

Analysis and Machine Intelligence, IEEE Transacti-

ons on, 33(9):1728–1743.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

386