A Hierarchical Approach for Indoor Action Recognition from New

Infrared Sensor Preserving Anonymity

elix Polla

, H

ene Laurent

and Bruno Emile

University of Orleans, Prisme Laboratory EA 4229, Orl

eans, France

INSA CVL, University of Orleans, Prisme Laboratory EA 4229, Bourges, France

Keywords:

Low Resolution Infrared Sensor, Motion History Image (MHI), Feature Selection, Action Recognition.

Abstract:

This article is made in the context of action recognition from infrared video footage for indoor installations.

The sensor we use has some peculiarities that make the acquired images very different from those of the

visible imagery. It is developed within the CoCAPS project in which our work takes place. In this context,

we propose a hierarchical model that takes an image set as input, segments it, constructs the corresponding

motion history image (MHI), extracts and selects characteristics that are then used by three classiﬁers for

activity recognition purposes. The proposed model presents promising results, notably compared to other

models extracted from deep learning literature. The dataset, designed for the CoCAPS project in collaboration

with industrial partners, targets ofﬁce situations. Seven action classes are concerned, namely: no action,

restlessness, sitting down, standing up, turning on a seat, slow walking, fast walking.

1 INTRODUCTION

In recent years, automatic recognition of human ac-

tivities has attracted a lot of attention in the ﬁeld

of computer vision. Action recognition can be used

for analyzing the behavior of people or for monitor-

ing living and workplace environments (Jalal et al.,

2017; Laptev et al., 2008). Despite impressive perfor-

mances, the main pitfall faced by the techniques de-

veloped for home automation is a marked restraint of

the users to be ﬁlmed. This led to feasibilty studies for

human activity recognition through unobtrusive sen-

sors such as pyroelectric infrared sensors (Luo et al.,

2017), bluetooth low-energy beacons (Filippoupoli-

tis et al., 2017) or acceleration sensors incorporated

in smartphones or smartwatches (Sefen et al., 2016).

To cope with the problems of personal identity re-

vealed in operation, one of the objectives of the Co-

CAPS project is to consider the feasibility of action

recognition using images coming from a low resolu-

tion (64 ×64 pixels) infrared sensor which guarantees

the respect of the intimacy of the person. The sen-

sor developed by Irlynx

within the project is a pro-

totype able to return images of moving objects in a

room. The technology is based on the principle of py-

roelectric detection, that is to say, the detection of dis-

www.irlynx.com

placement of a hot body present in the monitored vol-

ume. It allows to observe bright or low-light scenes

that would be otherwise difﬁcult to monitor. This type

of sensor, up-and-coming, is not yet completely mas-

tered. The main disadvantage is that the provided im-

ages present a very noisy aspect, which brings up a

problem of image quality (see Figure 1). Arranged in

the ceiling upon request from building industrialists,

we have then to face two difﬁculties: the considered

sensor provides unusual data and unusual ﬁeld of vi-

sion, so that no public database nor work dealing with

this problem are available.

In this article, we detail the database constructed

for the CoCAPS project, the proposed hierarchical ap-

proach for action recognition and its performances

that achieve competitiveness with models extracted

from deep learning literature.

2 DATA REPRESENTATION AND

FEATURE EXTRACTION

Activity recognition often breaks down into three

main phases: segmentation, data representation and

feature extraction, and ﬁnally classiﬁcation. Repre-

sentation and feature extraction have crucial inﬂuence

on the performance of recognition, therefore it is es-

Polla, F., Laurent, H. and Emile, B.

A Hierarchical Approach for Indoor Action Recognition from New Infrared Sensor Preserving Anonymity.

DOI: 10.5220/0008942002290236

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

229-236

ISBN: 978-989-758-402-2; ISSN: 2184-4321

229

Figure 1: The ﬁrst row presents images with visible camera respectively without and with the presence of a moving person in

the scene; the second row presents the corresponding images from the infrared sensor.

sential to extract or represent features of image frames

in a proper way. In this article, we describe some ap-

proaches used to represent a video sequence in order

to combine them and extract representative informa-

tion for each video.

2.1 Representation of Video Sequence

In 2D image sequences, many representations are

possible such as silhouettes (Bobick and Davis, 2001;

Gorelick et al., 2007), key pose (Liu et al., 2013), op-

tical ﬂow (Robertson and Reid, 2005), trajectory de-

scription (Wang et al., 2013), local descriptors: SIFT

(Lowe, 2004), HOG (Dalal and Triggs, 2005).

Given the main constraints related to the used sen-

sor and its usage context which are: noisy images, no

detection on images if there is no movement of a hot

body and sensor positioning in top view, we opt for an

approach based on the silhouette. This approach im-

plies that the segmentation and post-processing step

is effective so that the subsequent processing can op-

erate. In the work carried out previously (Polla et al.,

2017), a method of segmentation and post-processing

adapted to the sensor has been proposed.

Among the silhouette-based methods, some au-

thors proposed the concept of Spatio-temporal tem-

plate (Bobick and Davis, 2001). They ﬁrst segment

the images to extract the motion image, then for a set

of images, they construct the motion history image

(MHI) and the motion energy image. The MHI is a

temporal model of video representation quite simple

and widely used for the recognition of action (Ahad

et al., 2012). From this information, it is possible to

extract a large number of features.

In our approach, we use equation (1) to build the

motion history image. It allows to group consecutive

images of each video sequence according to their time

information.

MHI(x, y,t) =

max

{

1,..,m

}

{

D(x, y, t −(m − i)) − 10(m − i)

}

(1)

where D(x, y,t) is the binary motion detection im-

age extracted from the segmentation step at time t,

m represents the selected time frame (m=15) and the

value 10 is a threshold for varying the grayscale in the

construction of the motion history.

2.2 Feature Extraction

In the literature, feature extraction approaches are

classiﬁed into two families, namely contour-based

methods and region-based methods (see Figure 2).

We propose a framework combining the contour-

based method and the region-based method. In this

section, we brieﬂy present the three used descriptors.

2.2.1 Hu Moments

In this approach, central moments of any order are

computed from raw moments and standardized to

construct an invariant descriptor in translation and

in scale. Finally, to also achieve invariance to the

rotation, Hu (Hu, 1962) reformulated the above-

mentioned moments by deﬁning 7 new measures.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

230

Figure 2: Numerous approaches for shape representations.

2.2.2 Color Histogram of Oriented Phase

(CHOP)

The CHOP descriptor (Ragb and Asari, 2016) can ac-

curately identify and locate image characteristics on

gradient-based techniques. The features are formed

by extracting the phase congruence information for

each pixel, convolving the image with a pair of

quadrature log-Gabor ﬁlters to extract the local fre-

quencies and phase information.

2.2.3 Geometrical Shape

Several geometric indices have been proposed in the

literature (Coster and Chermant, 1989). Some exam-

ples used for this study are : area (A), ellipse charac-

teristics (eccentricity, major axis: R

max

, minor axis:

min

), convex hull (C

), perimeter (P), circularity

(

min

max

), perimeter convexity (

P(C

)

), surface convex-

ity (

A(C

)

2.3 Feature Selection

On a single image, a large number of data can be ex-

tracted. This can result in the known problem of the

”dimension curse” related to classiﬁcation algorithms

because considering a high number of attributes in-

creases the risk of taking into account redundant or

correlated ones which makes these algorithms more

complex (storage space and high learning time) and

sometimes less effective. In the literature, the feature

selection approaches are divided into 2 categories:

feature ranking algorithms (Biesiada et al., 2005) and

subset selection algorithms (Yu and Liu, 2003; Hall,

1999). In this article we use 3 methods of feature se-

lection from the literature to ﬁnd the more adapted

one for the considered situation.

CFS (Correlation-based Feature Selection) (Hall,

1999) is an algorithm that classiﬁes subsets of enti-

ties according to a heuristic evaluation function based

on correlation. The bias of the evaluation function is

oriented towards subsets containing entities strongly

correlated with the class and having low correlation

between them. Non-relevant entities should be ig-

nored as they have a low correlation with the class.

Redundant features must be hidden because they will

be highly correlated with one or more of the remain-

ing features.

ILFS (Inﬁnite Latent Feature Selection) (Roffo

et al., 2017) is a robust probabilistic feature selection

algorithm based on a probabilistic latent graph that

performs the ranking step while considering all pos-

sible subsets of features, such as paths on the graph.

The relevance of an attribute is modeled as a latent

variable in a PLSA-inspired generation process (Hof-

mann, 1999) that allows the importance of a charac-

teristic to be studied when it is added to a set of at-

tributes.

MRMR (Minimum Redundancy and Maximum

Relevance) (Peng et al., 2005) is a feature selection

method that allows to select features that are mutually

distant from each other while having a high correla-

tion with the classiﬁcation variable. A comparative

study on measures to assess the best redundancy and

relevance was presented in (Auffarth et al., 2010).

3 OVERVIEW OF THE

PROPOSED METHOD

For the CoCAPS project, several case studies have

been targeted for action recognition. One of which

relates to ofﬁce environments, another one to insti-

tutions for older people. In these contexts manufac-

turers are interested in various scenarios that lead to

A Hierarchical Approach for Indoor Action Recognition from New Infrared Sensor Preserving Anonymity

231

Figure 3: Overview of the proposed approach.

the deﬁnition of fast walk / slow walk, big and small

movements.

A movement is considered to be a big movement

if there is a displacement of the center of gravity of

the person and is considered as a small movement in

the opposite case. Then in each group we ﬁnd sev-

eral actions. For example, in big movement we have

fast walk (> 1 meter/second) and slow walk (< 1 me-

ter/second). In small movement we have actions like:

to sit down or get up, to stir (which includes for ex-

ample typing, answering the phone, browsing docu-

ments) and to turn on a seat.

We propose a hierarchical model that exploits the

Motion History Image and extracts a set of features

from the MHI including geometric shapes, Hu mo-

ments and CHOP. In our approach, we deﬁne 3 classi-

ﬁers: the ﬁrst one to separate no movement, big move-

ment and small one, the second one to separate the

actions related to big movement and the last one for

actions related to small movement.

Figure 3 shows an overview of the method. For

each classiﬁer, the deﬁnition of a dedicated features

set is essential to have good performance. The left

part of Figure 3 corresponds to the training phase of

the model, within which we tested different feature

sets including different combinations. During the test

phase, we do not make a selection of features, but ex-

ploit the descriptors identiﬁed as being the best.

4 EXPERIMENTATION AND

RESULTS

4.1 Dataset

We perform our experiments on the data acquired for

the CoCAPS project. The dataset shows common ac-

tions in ofﬁces. It has 7 classes of human action:

no action, restlessness, sitting, standing, turning on

a seat, slow walking (speed less than 1 meter/second)

and fast walking (speed greater than 1 meter/second).

In total, the dataset consists of 700 videos samples

(100 samples per action). The videos were taken con-

sidering various players and were collected at differ-

ent times of the year. We consider clip sizes of 15

frames (about 1.5 second) for action recognition. In

Figure 4, we present some results of the motion his-

tory image obtained for each action.

4.2 Protocol of Validation

We propose to conduct the performance study while

using KNN (k-nearest neighbors) classiﬁcation which

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

232

Figure 4: Example of motion history images for no action

(a), restlessness (b), sitting or standing (c), turning on seat

(d), slow walking (e) and fast walking (f).

is commonly used in machine learning. KNN is a

classiﬁcation method based on the closest training

samples. To estimate the class associated with a new

input, KNN algorithm consists in taking into account

(in an identical way) the k learning samples which

are the closest to the new input, according to a dis-

tance to be deﬁned. There are different methods for

comparing these values, like Hamming distance, Ma-

halanobis distance, Euclidean distance, etc. After var-

ious testings we used for our model the Manhattan

distance, also known as city block distance, and k=3

neighbors. The Manhattan distance is calculated by

making the sum of absolute differences between the

coordinates of a pair of objects. This distance pro-

duces results close to those obtained by the simple

Euclidean distance. However, with this measure, the

effect of only one signiﬁcant difference (outlier) is at-

tenuated (Mohibullah et al., 2015).

To validate our results we use the K-fold valida-

tion (K=10) because it avoids the over-learning of the

designed model. The K-Fold or cross-validation is

a protocol in which individuals are separated into K

groups of identical sizes. The learning takes place

on K-1 groups and the validation on the withdrawn

group. This operation is repeated for all groups and

the average recognition rate is calculated. For model

accuracy measure we use the F-score (Equation (2)).

F-score = 2.

precision.recall

precision + recall

(2)

where precision is the number of true positive results

divided by the number of all positive results returned

by the classiﬁer and recall is the number of correct

positive results divided by the number of all relevant

samples.

4.3 First Results and Interpretations

Before presenting the different results of the proposed

hierarchical model, we ﬁrst present those of the sim-

pliest model (see Table 1) consisting in the raw clas-

siﬁcation into 7 classes (no action, restlessness, sit-

ting, standing, turning on a seat, slow walking and

fast walking) obtained with each type of descriptors

or each combination of descriptors. The values of the

f-scores do not exceed 83%, achieved when we just

use the CHOP descriptor as features. Table 2 shows

the confusion matrix obtained in that most favourable

case. One can note that slow walking, within big

movement, is not always appropriately recognized. It

is mainly confused with fast walking and, to a lesser

extent, with restlesness and sitting. Sitting is also con-

fused with standing and inversely. As the classes of

small and big movements are not very similar and

also because it meets the needs of industrialists, we

decided to test a hierarchical classiﬁcation approach

to ﬁrst differentiate between no/big/small movement

(classiﬁer 1) and then classify within big movement

(classiﬁer 2) and small movement (classiﬁer 3).

Table 1: Values of F-score for a model with a raw classiﬁ-

cation into 7 classes (single classiﬁer).

Features

KNN

(7 classes)

Hu (7 features) 49%

Geo (9 features) 64%

CHOP (128 features) 83%

Geo+Hu (16 features) 50%

Hu+CHOP (135 features) 71%

Geo+CHOP (137 features) 65%

Geo+Hu+CHOP (144 features) 52%

Table 2: Confusion matrix with CHOP as features: F-

score=83%.

We present in Table 3, the F-score values of the

3 classiﬁers with all combinations of the descriptors

presented above. The ﬁrst observation is that the sep-

aration of no, big and small movement (classiﬁer 1)

is easily done by using the geometrical shape descrip-

tors or the combination of geometrical shape descrip-

tors and CHOP with 97.8% of F-score.

The results of Table 3 also show that by merg-

ing descriptors, one can increase the performance of

a classiﬁer. This is the case of classiﬁer 2 (big move-

ment) and classiﬁer 3 (small movement) where the fu-

sion of CHOP and Hu moment descriptors allows to

achieve 93.5% of F-score for classiﬁer 2 and 82.5 %

of F-score for classiﬁer 3.

We tested the proposed hierarchical approach us-

A Hierarchical Approach for Indoor Action Recognition from New Infrared Sensor Preserving Anonymity

233

Table 3: Classiﬁcation rate of 3 classiﬁers with KNN.

Features

no/big/small movements big movements small movements

classiﬁer 1 classiﬁer 2 classiﬁer 3

Hu (7 features) 68.8% 90.5% 61%

Geo (9 features) 97.8% 86.5% 47.7%

CHOP (128 features) 97.1% 89.5% 81.5%

Geo+Hu (16 features) 84.4% 86.5% 47.7%

Hu+CHOP (135 features) 83.5% 93.5% 82.5%

Geo+CHOP (137 features) 97.8% 87% 51.5%

Geo+Hu+CHOP (144 features) 84.4% 87% 51.7%

ing each time the most relevant descriptors for build-

ing each classiﬁer. Two approaches were identiﬁed:

• Approach 1: using geometrical shape descriptors

for classiﬁer 1 and Hu moments + CHOP for clas-

siﬁer 2 and 3.

• Approach 2: using geometrical shape descriptors

+ CHOP for classiﬁer 1 and Hu moments + CHOP

for classiﬁers 2 and 3.

The hierarchical approach allows to achieve a F-score

value of 86.5% for the ﬁrst approach and 86.7% for

the second approach.

Table 4 presents the confusion matrix of the pro-

posed hierarchical approach 2. One can note that

the recognition rate of quite all classes has been im-

proved. The results are more uniform, especially for

small movement. However confusions remain mostly

between slow/fast walking, between sitting/standing

and between turning on a seat/restlessness.

Table 4: Confusion matrix with hierarchical approach: F-

score=86.7%.

4.4 Results of Feature Selection

While combining different types of descriptors may

improve the classiﬁcation rate, having a large number

of features for the model may conversely reduce its

performance (calculation time and classiﬁcation per-

centage). Ensuring a balance is necessary and the

question of the choice of relevant features has to be

dealt with. That is why we considered feature selec-

tions using three relevant methods from the literature,

and analyzed the impact of the number of features for

each method. These tests will provide the best fea-

tures to constitute the classiﬁers.

To separate no, big and small movements, we use

the combination of geometrical shape features and

CHOP. In Figure 5 we note that by using the ﬁrst 60

descriptors provided by the ILFS and MRMR meth-

ods, the F-score value is slightly increased.

Figure 5: Results of feature selections for classiﬁer 1: no,

big, small movements.

Figure 6: Results of feature selections for classiﬁer 2: big

movements.

Figure 6 shows that for classiﬁer 2, by applying

the MRMR selection method on the 135 descriptors

resulting from the combination of Hu moments and

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

234

CHOP and using only the ﬁrst 30 selected features,

we can gain more than 2% of performance.

Concerning the selection of features for classiﬁer

3 (small movement), Figure 7 shows that instead of

the 135 descriptors resulting from the combination of

Hu and CHOP, we can use 120 features selected by

ILFS method while achieving a slightly increased per-

formance.

Figure 7: Results of feature selections for classiﬁer 3: small

movements.

Table 5 presents the confusion matrix of the pro-

posed hierarchical approach using for classiﬁer 1, the

ﬁrst 60 descriptors provided by the ILFS method, then

for classiﬁer 2, the ﬁrst 30 descriptors provided by

MRMR, and ﬁnally for classiﬁer 3, the ﬁrst 120 de-

scriptors provided by the ILFS method. This selection

allows to achieve a F-score value of 87.5%. It leads

to an homogenization of the recognition rates within

small movement but still fails discriminating slow and

fast walking.

Table 5: Confusion matrix with feature selection: F-

score=87.5%.

4.5 Evaluation Review

We also compared the performance of the pro-

posed hierarchical model with one depth learning ap-

proach, namely 3D-CNN (3D-Convolutional Neural

Network), which is widely used for action recognition

(Ji et al., 2012). In (Polla et al., 2019) it is shown that

the 3D-CNN model is more efﬁcient for our database

than the LSTM model or 3D-CNN combined with the

LSTM. This method aims at learning motion features

by learning a hierarchy consisting of multiple lay-

ers of 3D spatio-temporal convolution kernels whose

last layer output is used by a Multi Layer Perceptron

(MLP) for classiﬁcation.

Table 6: Summary table of comparisons.

Models

F-score

(7 classes)

best result for raw classiﬁcation 83%

3D-CNN with MLP 85%

hierarchical without selection 86.7%

hierarchical with selection 87.5%

Table 6 summarizes the different classiﬁcation re-

sults obtained for the different tests. The deep learn-

ing model only achieve a F-score of 85%, which

stresses the relevance of the proposed model.

5 CONCLUSION AND

PERSPECTIVES

In this article, we present a hierarchical model for ac-

tion recognition using low-resolution infrared video.

This approach results in a classiﬁcation rate of 87.5%

while with a raw classiﬁcation approach, 83% of F-

score is achieved. The proposed approach even ex-

ceeds advanced methods such as the one based on

neural networks. These results are quite interesting

considering the type and quality of the sensor, con-

straint of the project (position of sensor in top view)

and also the non-similarity of the shape within a same

class (see Figure 4.(b) restlessness, for example). In

future work, we plan to do action recognition on long

videos. This will allow to answer the problematic of

video sequence clipping for the classiﬁcation.

ACKNOWLEDGEMENTS

The authors wish to thank BPI France, the Regional

Councils of Limousin and Rh

one-Alpes with FEDER,

Department Council of Is

ere and District town of

Bourges, Bourges Plus, for providing the ﬁnancial

support to CoCAPs project. The CoCAPs project,

from FUI N

◦

20, is also supported by clusters S2E2,

Minalogic.

A Hierarchical Approach for Indoor Action Recognition from New Infrared Sensor Preserving Anonymity

235

REFERENCES

Ahad, M. A. R., Tan, J. K., Kim, H., and Ishikawa, S.

(2012). Motion history image: its variants and appli-

cations. Machine Vision and Applications, 23(2):255–

281.

Auffarth, B., L

opez, M., and Cerquides, J. (2010). Compar-

ison of redundancy and relevance measures for feature

selection in tissue classiﬁcation of ct images. In In-

dustrial Conference on Data Mining, pages 248–262.

Springer.

Biesiada, J., Duch, W., Kachel, A., Maczka, K., and

Palucha, S. (2005). Feature ranking methods based on

information entropy with parzen windows. In Interna-

tional Conference on Research in Electrotechnology

and Applied Informatics, pages 1–9.

Bobick, A. F. and Davis, J. W. (2001). The recognition

of human movement using temporal templates. IEEE

Transactions on Pattern Analysis & Machine Intelli-

gence, (3):257–267.

Coster, M. and Chermant, J.-L. (1989). Pr

ecis d’analyse

d’images. Technical report, Presses du CNRS.

Dalal, N. and Triggs, B. (2005). Histograms of oriented

gradients for human detection. In Computer Vision

and Pattern Recognition (CVPR).

Filippoupolitis, A., Oliff, W., Takand, B., and Loukas, G.

(2017). Location-enhanced activity recognition in

indoor environments using off the shelf smartwatch

technology and ble beacons. Sensors, 17(6):1230.

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

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

transactions on pattern analysis and machine intelli-

gence, 29(12):2247–2253.

Hall, M. A. (1999). Correlation-based feature selection for

machine learning. PhD thesis, New Zealand, Depart-

ment of Computer Science, Waikato University.

Hofmann, T. (1999). Probabilistic latent semantic analysis.

In Proceedings of the Fifteenth conference on Uncer-

tainty in artiﬁcial intelligence, pages 289–296.

Hu, M.-K. (1962). Visual pattern recognition by moment

invariants. IRE transactions on information theory,

8(2):179–187.

Jalal, A., Kim, Y.-H., Kim, Y.-J., Kamal, S., and Kim, D.

(2017). Robust human activity recognition from depth

video using spatiotemporal multi-fused features. Pat-

tern recognition, 61:295–308.

Ji, S., Xu, W., Yang, M., and Yu, K. (2012). 3d convolu-

tional neural networks for human action recognition.

IEEE transactions on pattern analysis and machine

intelligence, 35(1):221–231.

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

B. (2008). Learning realistic human actions from

movies.

Liu, L., Shao, L., and Rockett, P. (2013). Boosted

key-frame selection and correlated pyramidal motion-

feature representation for human action recognition.

Pattern recognition, 46(7):1810–1818.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60(2):91–110.

Luo, X., Guan, Q., Tan, H., Gao, L., Wang, Z., and Luo,

X. (2017). Simultaneous indoor tracking and activity

recognition using pyroelectric infrared sensors. Sen-

sors (Basel), 17(8):1738.

Mohibullah, M., Hossain, M. Z., and Hasan, M. (2015).

Comparison of euclidean distance function and man-

hattan distance function using k-mediods. Interna-

tional Journal of Computer Science and Information

Security, 13(10):61.

Peng, H., Long, F., and Ding, C. (2005). Feature selec-

tion based on mutual information: criteria of max-

dependency, max-relevance, and min-redundancy.

IEEE Transactions on Pattern Analysis & Machine In-

telligence, (8):1226–1238.

Polla, F., Boudjelaba, K., Emile, B., and Laurent, H.

(2017). Proposal of segmentation method adapted

to the infrared sensor. In International Conference

on Advanced Concepts for Intelligent Vision Systems

(ACIVS), pages 639–650. Springer.

Polla, F., Laurent, H., and Emile, B. (2019). Action recog-

nition from low-resolution infrared sensor for indoor

use: a comparative study between deep learning and

classical approaches. In 20th IEEE International Con-

ference on Mobile Data Management (MDM), pages

409–414.

Ragb, H. K. and Asari, V. K. (2016). Color and local

phase based descriptor for human detection. In 2016

IEEE National Aerospace and Electronics Conference

(NAECON) and Ohio Innovation Summit (OIS), pages

68–73.

Robertson, N. and Reid, I. (2005). Behaviour understand-

ing in video: a combined method. In Tenth IEEE In-

ternational Conference on Computer Vision (ICCV),

volume 1, pages 808–815.

Roffo, G., Melzi, S., Castellani, U., and Vinciarelli, A.

(2017). Inﬁnite latent feature selection: A probabilis-

tic latent graph-based ranking approach. In Proceed-

ings of the IEEE International Conference on Com-

puter Vision, pages 1398–1406.

Sefen, B., Baumbach, S., Dengel, A., and Abdennadher, S.

(2016). Human activity recognition using sensor data

of smartphones and smartwatches. In Proceedings of

the 8th International Conference on Agents and Artiﬁ-

cial Intelligence (ICAART), Volume 2, pages 488–493.

Wang, H., Kl

aser, A., Schmid, C., and Liu, C.-L. (2013).

Dense trajectories and motion boundary descriptors

for action recognition. International journal of com-

puter vision, 103(1):60–79.

Yu, L. and Liu, H. (2003). Feature selection for high-

dimensional data: A fast correlation-based ﬁlter so-

lution. In Proceedings of the 20th international con-

ference on machine learning (ICML), pages 856–863.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

236