Activity Recognition based on High-Level Reasoning

An Experimental Study Evaluating Proximity to Objects and Pose Information

Julia Richter, Christian Wiede, Enes Dayangac, Markus Heß and Gangolf Hirtz

Department of Electrical Engineering and Information Technology

Chemnitz University of Technology, Reichenhainer Str. 70, 09126 Chemnitz, Germany

Keywords:

Video Analysis, 3-D Image Processing, Activity Recognition, Pose Estimation, High-Level Reasoning,

Ambient Assisted Living.

Abstract:

In the context of Ambient Assisted Living (AAL), the detection of daily activities is an active ﬁeld of research.

In this study, we present an algorithm for the performed Activities of Daily Living (ADLs) related to personal

hygiene, which is based on the evaluation of a person’s proximity to objects and pose information. To this end,

we have employed a person detection algorithm that provides a person’s position within a room. By fusing the

obtained position with the objects’ position, we were able to deduce whether the person was occupied with a

certain object and to draw conclusions about the performed ADLs. One prerequisite for a reliable modelling

of human activities is the knowledge about the accuracy of the person detection algorithm. We have, therefore,

analysed the algorithm with regard to its accuracy under different, application-speciﬁc conditions. The results

show that the considered algorithm ensures high accuracy for our AAL application and that it is even suitable

for environments, in which objects are very close to each other. On the basis of these ﬁndings, tests with

video sequences have been conducted in an AAL environment. This evaluation conﬁrmed that the reasoning

algorithm can reliably recognise activities related to personal hygiene.

1 INTRODUCTION

In recent years, a lot of studies have been set on

the development of technical support systems for the

maintenance of human care standards. The central

aim of the AAL system we have developed together

with our medical partners is to provide assistance for

elderly people and their caregivers, so that the elderly

can continue to live in their familiar environment in-

stead of moving to a nursing home. Besides assisting

elderly, our system supplies caregivers and relatives

with information that helps to optimise the care pro-

cess. In this study, we focus on elderly people at an

early stage of dementia, ﬁrstly because they are the

group most likely to forget their daily routines, and

secondly because they are often incapable of commu-

nicating to their caregivers, which daily routines they

have already performed. We aim, therefore, at pro-

viding caregivers with meta information about their

patients’ daily activities (ADLs). Based on this infor-

mation, caregivers can be supported in planning the

individual care process and in assessing their patients’

need of care.

For this purpose, we detect ADLs by utilising so-

called smart sensors. These sensors consist of stereo

cameras with an internal processing unit that monitors

ADLs. We thus attach high importance to privacy is-

sues: the sensors, which are mounted at the ceiling

of the living environment, make information about

ADLs available – without releasing raw image data,

as it is stored in a database in the form of meta data.

This system also generates reminding messages for

the elderly if they failed to perform certain activities.

Furthermore, the contents of the database can be ac-

cessed by the caring personnel via a web interface. By

this means, the caring personnel can obtain relevant

information that has been inaccessible so far. It en-

ables caregivers to interpret a patient’s uncooperative

behaviour in the morning when a disturbed sleeping

pattern has been detected the previous night. More-

over, drastic changes in the daily activities can be no-

ticed, so that caregivers can react promptly and adapt

the caring process to the actual situation. These ex-

amples illustrate how the assistance and information

system can contribute to a better understanding of the

patient’s behaviour and improve the caring process.

With the presented approach, it will be possible for

caregivers to respond appropriately to their patients’

individual needs. Current assistance systems have not

yet considered the above mentioned aspects.

Richter, J., Wiede, C., Dayangac, E., Heß, M. and Hirtz, G.

Activity Recognition based on High-Level Reasoning - An Experimental Study Evaluating Proximity to Objects and Pose Information.

DOI: 10.5220/0005658804150422

In Proceedings of the 5th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2016), pages 415-422

ISBN: 978-989-758-173-1

415

2 RELATED WORK

In order to identify special demands or detect emer-

gencies, several approaches have been developed that

monitor a person’s daily activities, for example the

use of motion detectors (Steen et al., 2013) or body-

worn sensors (Scanaill et al., 2006). In the work of

Pirsiavash et al. (Pirsiavash and Ramanan, 2012),

ADLs are detected by a wearable camera that pro-

cesses the ﬁrst-person camera view. In our project,

we have decided to employ non-wearable, optical sen-

sors, i. e. wide-angle stereo cameras, since people

with dementia are apt to put wearable sensors off and

to forget them. Besides, more valuable information

can be derived from image data than from motion de-

tectors.

One indicator for ADLs is the room in which the

person stays (Steen et al., 2013; Richter et al., 2014).

In the context of an AAL project, Richter et al. have

introduced a person detection algorithm that assigns

a room to the person’s current position. Based on

the chronological order of the rooms a person en-

tered, conclusions about the performed ADLs could

be drawn. If the person was detected in the sleeping

room at the beginning of the day and afterwards in

the bathroom for a few minutes, it can be inferred that

the person has attended to his or her personal hygiene.

Conclusions about other ADLs, such as the prepara-

tion of food, which is usually done in the kitchen, or

the leaving of the ﬂat, can be drawn in a similar way.

However, for a more reliable ADL prediction, the

analysis of the room alone is insufﬁcient. For this rea-

son, we have developed an algorithm that deduces the

performed activities by evaluating the proximity of a

person to certain objects in the room. Moreover, pose

information obtained by a machine learning based ap-

proach (Richter et al., 2015) is used in order to draw

conclusions about the activity. The presented study

focuses on three objects in the bathroom, the sink, the

shower and the toilet. This choice was mainly moti-

vated by medical reasons. The frequency of toileting,

for example, provides relevant information for diag-

nosing and treating incontinence.

In order to determine the proximity of a person

to a certain object, the person’s position in the room

has to be known. Due to the large number of algo-

rithms, a brief summary of the state-of-the-art person

detection algorithms should sufﬁce here. Harville et

al. (Harville and Li, 2004) used a stereo camera and a

person detection algorithm similar to the one already

described (Richter et al., 2014). By using the gener-

ated depth map as an input, they estimated the virtual

overhead view (plan view), which they used for per-

son detection and tracking. Furthermore, they dealt

with occlusions and 3-D noise by combining occu-

pancy and height maps. Yous et al. (Yous et al.,

2008) introduced the world z-map, which represents

the z coordinate of the corresponding world point

for every pixel. Their algorithm utilises the fact that

the boundaries between close objects persist, whereas

they merge on the plane view. In this approach, the

height of the mounted camera in relation to the ﬂoor

and the angle in relation to the wall are required for

the determination of the world coordinates of a certain

point.

Both these studies have revealed shortcomings of

their stereo vision-based algorithms. Harville et al.

(Harville and Li, 2004) determined a point-wise mean

positional error from ground truth of 16 cm. Yous et

al. (Yous et al., 2008) evaluated their algorithm by a

video in which each detected person was marked by a

cuboid. Furthermore, false positive and false negative

rates were calculated. However, neither of the studies

provided detailed numerical results about spatial ac-

curacy. When applying a person detection algorithm,

it is essential to know the grade of reliability of the de-

termined position in order to draw conclusions about

the performed activity. If the person is detected close

to the toilet, for example, it is necessary to know how

reliable this information is in order to avoid misin-

terpretations. If the error of the person detection al-

gorithm reaches or exceeds the distance between the

objects, the assignment to a certain object is unreli-

able.

In our project, we utilise the person detection al-

gorithm described in (Richter et al., 2014) with the

aim of reﬁning the room assignment to the assignment

of objects the person is probably occupied with. Since

detailed spatial accuracy measurements have not yet

been considered for this algorithm and since there is

no standard method to be found in literature, we have

developed a method to determine the accuracy in dif-

ferent conditions and scenarios (walking speed, walk-

ing direction, exposure time, distance from sensor).

By means of this analysis, the accuracy of the algo-

rithm described by Richter et al. (Richter et al., 2014)

has been determined in order specify the occurring

errors and that might lead to false conclusions with

regard to the activities. We thus introduce a method,

which can also be applied to other person detection

algorithms that determine a person’s position in the

world.

The accuracy analysis presented in this study

yields that persons can be localised in a very accu-

rate way, so that the determined position can be relied

on even if the speciﬁc objects are very close to each

other. On the basis of this ﬁnding, we have developed

a reasoning algorithm that recognises activities per-

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

416

formed in a bathroom. In addition, we have evaluated

the algorithm in our testing ﬂat under realistic con-

ditions by comparing the activity determined by the

algorithm with labelled video sequences.

The algorithm for reasoning about ADLs is pre-

sented in section 3, whereas the employed and anal-

ysed person detection algorithm and the pose esti-

mation algorithm are brieﬂy summarized in section

3.1 and section 3.2 respectively. Section 4 presents

the methods of analysis for the person detection and

the reasoning algorithm. The results of both analyses

are presented and discussed in sections 5.1 and 5.2.

In section 6, conclusions about the accuracy of both

the person detection and the reasoning algorithm are

drawn. Besides, an outline regarding the detection of

further ADLs by employing the presented approach is

given.

3 METHODS AND

IMPLEMENTATION

3.1 Location Data Acquisition from

Person Detection Algorithm

The employed person detection algorithm (Richter

et al., 2014) is able to locate several persons in 3-

D coordinates by processing the 3-D point cloud de-

rived from stereo data. In principal, this algorithm

works for any sensor providing a 3-D point cloud. In

our study, high-quality wide ﬁeld of view lenses with

a focal length of 3.5 mm and small radial distortion

are employed. The image resolution is 1360 × 1024

pixels. The stereo sensor has a baseline distance of

150 mm. Firstly, the algorithm derives foreground hy-

potheses from the world z-map by applying a mixture

of Gaussian segmentation method (Zivkovic, 2004),

while the learning rate is set to 0.01 in our study. All

points belonging to the foreground are projected onto

the ﬂoor plane that was previously deﬁned during ex-

trinsic camera calibration. In this ﬂoor plane view,

blobs of a certain size are detected. The maximum

z value in a blob deﬁnes the height, the size of the

blob deﬁnes the width as well as the length of the

cuboid that ﬁnally characterises the detected person.

The center of the cuboid denotes the person’s loca-

tion in the scene. In Figure 1, a detected person of

an example recording is shown. At this point, the x

and y components of the center point, i. e. the pro-

jection of the location onto the ﬂoor p

= (x

), are

used for the proximity determination, whereas the z

component is not relevant at this point.

During the accuracy analysis, the person is al-

ways moving and therefore constantly detected as

foreground. In our practical AAL application, how-

ever, when a person is resting for a longer time, previ-

ously detected foreground pixels will become back-

ground. In order to localise persons even in such

cases, we additionally detect persons by employing

a head-shoulder detection algorithm (Dayangac et al.,

2015).

Figure 1: Example point cloud of a proband in our testing

ﬂat. The detected person is characterised via a cuboid in

3-D data.

3.2 Pose Information

Information about the general pose is obtained by us-

ing the algorithm described by (Richter et al., 2015).

In this approach, a classiﬁer is trained using a Support

Vector Machine by assigning the points belonging to

the person’s point cloud according to their z compo-

nent to vertically aligned bins. Depending on the cal-

culated feature vector, the classiﬁer predicts whether

the person is standing, sitting or lying. In our study,

we evaluate whether a person is sitting or standing.

3.3 Reasoning About ADLs

For reasoning about ADLs, it is required to determine

whether the detected person is close to a certain ob-

ject. Therefore, the stereo sensors that are distributed

in the test ﬂat have been calibrated in such a way

that they all share the same world coordinate system.

The objects’ positions (objects’ centres) with respect

to the origin of the coordinate system as well as the

expansions of relevant and ﬁxed object, such as bed,

refrigerator, shower, basis and toilet, have been stored

in a look-up-table (LUT). Consequently, if there are n

objects, the LUT contains n entries with the according

Activity Recognition based on High-Level Reasoning - An Experimental Study Evaluating Proximity to Objects and Pose Information

417

object centres and their expansions. These expansions

serve as thresholds thresh

for the proximity determi-

nation.

In order to determine whether a person is close to

an object, the distance dist

between the person’s po-

sition p

= (x

) and all the objects’ positions p

o,n

) is compared with the according expansion

values thres

in the LUT, whereas n = {1,2,..., N},

Hereby, N denotes the number of objects and n the

index of a speciﬁc object. If the distance between the

person’s center and the object’s center is smaller than

the deﬁned threshold value in the LUT, the person is

considered to be interacting with the object. In this

case, the boolean variable close is 1, otherwise it is

0. The following two equations are applied n times if

there are n objects in a room. In the presented project,

three objects, i. e. sink, shower and toilet, are present

in the bathroom of our testing ﬂat.

dist

o,n

− p

, (1)

close =

(

1, if dist

< thresh

0, otherwise.

(2)

At this point, it should be stated that the bathroom

is relatively small (approximately 2.00 m × 1.80 m)

compared to other rooms in our testing ﬂat. This

results in a more challenging assignment, because the

objects are very close to each other.

This procedure, in combination with pose infor-

mation (standing and sitting), allows us to reason

about the following ADLs:

- Activities that are typical performed when standing

in front of a sink, such as washing hands, comb-

ing, teeth brushing, etc., if the person is close to

the sink and standing.

- Using the toilet if the person is close to the toilet

and sitting.

- Taking a shower if the person is close to the shower

and standing.

If the person is detected in the shower and stand-

ing, then the algorithm outputs the activity ”Taking a

shower”. If the person is detected close to the sink,

then it is concluded that the person performs activi-

ties that are typically for standing in front of a sink,

like ”washing hands, combing, teeth brushing, etc.”.

Similarly, if the person’s center is near the toilet and

the person is detected to be sitting, we reason that the

person probably uses the toilet. If none of the previ-

ously mentioned scenarios occur, we assume that the

person is doing another action.

Figure 2 shows a scenario in the bathroom of our

testing ﬂat, where a person is attending to his or her

personal hygiene.

Figure 2: The person is detected very close to the toilet.

Moreover, the algorithm determined that the person is sit-

ting. Based on this information, the system reasons that the

person is probably using the toilet. This view was gener-

ated in debug mode. Due to privacy aspects, it is not visible

outside of the smart sensor in the ﬁnal application.

4 ANALYSIS METHODS

In the following section, the analysis methods of both

the person detection algorithm and the reasoning al-

gorithm for the ADLs are presented.

4.1 Person Detection

In order to determine the the accuracy of the person

detection algorithm (Richter et al., 2014), the output

of this algorithm is compared with the position infor-

mation provided by a reference system. Each time

both a stereo measurement p

and a reference mea-

surement p

are obtained, the error e of the stereo

measurement is calculated as the Euclidean distance

between these two values.

e =

− p

. (3)

Error histograms are generated for every scenario

on the basis of these error measurements. They illus-

trate the relative occurrence of the errors within dif-

ferent error intervals.

4.1.1 Position Data Acquisition from Reference

System

The reference system is completely independent from

the stereo sensor system. It is composed of infra-red

emitters and cameras that are able to detect special

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

418

rigid bodies. The reference system provides reliable

data at a frame rate of 120 FPS, so that higher walk-

ing speeds have no critical inﬂuence on the reference

measurement. In order to obtain reliable reference

data even with occurring reﬂections and occlusions,

eight of such devices are installed in the surveyed

volume, so that a good coverage with redundancy is

achieved. These emitters and cameras are distributed

along a rectangle on the ceiling that surrounds the sur-

veyed area. The rigid body that has been designed for

this accuracy analysis is a symmetric construction of

four infra-red reﬂecting markers. This device is in-

stalled in the center of a helmet, which a person wears

on the head during all recordings. This construction

is designed in such a way that it does not inﬂuence

the stereo measurement. The reference system out-

puts the 3-D center position of the rigid body. In this

study, only the x and the y coordinates of the 3-D po-

sition, denoted as p

= (x

), are used.

In order to directly compare the measured posi-

tion and the reference position, both the reference and

the stereo camera system are extrinsicly calibrated, so

that they share the same world coordinate system.

4.1.2 Analysis Method

In order to obtain statements about the accuracy, the

following different parameter conﬁgurations are in-

vestigated:

Table 1: Description of the scenario conﬁgurations.

Scenario Description

1 Distance from the stereo sensor, de-

termined for the scenario conﬁgura-

tions 2 - 3

2 Moving speed: slow (≈ 0.3 m/s) vs.

fast (≈ 1.3 m/s)

3 Moving direction: parallel vs. per-

pendicular to the optical axis of the

stereo sensor

4 Exposure time: high (40 ms) vs.

low (10 ms)

By setting the mentioned exposure times, a dark

scene with low background illumination as well as a

bright scene with very high illumination have been re-

produced. For scenario 2 to scenario 4, two video se-

quences with one person walking within the surveyed

area have been recorded for each scenario, whereas

the exposure time was 20 ms for scenario 2 and 3. All

these recordings were used for scenario 1.

Moreover, the mean, minimum and maximum er-

ror have been calculated for every scenario. In order

to investigate how the distance between sensor and

person inﬂuences the error, the mean error has been

calculated for different radius intervals using all the

recordings from scenario 2 to scenario 4. The inter-

val limits are 500 mm, 1000 mm, 1500 mm, 2000 mm,

2500 mm, 3000 mm and 3500 mm (see Figure 3). For

each interval, the mean error is calculated and plotted

in the middle of the interval.

4.2 Reasoning Algorithm

In order to evaluate the reasoning algorithm, three

video sequences with three different persons have

been recorded in the bathroom of our testing ﬂat. In

these sequences, a typical morning routine is repro-

duced: Each testing person is entering the bathroom

and attending to his or her usual personal hygiene.

Thereby, typical activities, such as using the toilet,

washing the hands and showering, are performed. Ev-

ery frame is recorded with a time-stamp. At the same

time, the system subsequently outputs the determined

ADL, accompanied by the time-stamp, into a ﬁle. Af-

ter the sequence has been recorded, every frame was

labelled with the actual ADL. Afterwards, both the

labelled ADL as well as the action that has been de-

termined by our system is plotted over time. In this

way, the output of the system can be compared against

ground-truth data. Moreover, it is possible to identify

potential delays the system exhibits.

5 RESULTS AND DISCUSSION

5.1 Person Detection

In this section, the results of the person detection anal-

ysis are presented and discussed for each scenario.

Figure 3 illustrates the inﬂuence of the distance

to the mean of the error, whereas the information be-

tween the measured mean values is linearly interpo-

lated.

The graphs show a higher error between the po-

sition measured by the stereo camera and the refer-

ence position in the range with a small distance to the

stereo sensor (≤ 1000 mm). For distances higher than

1000 mm, a tendency of an increasing mean error can

be observed. Generally, it can be stated that the mean

error will further increase at higher distance, because

the accuracy of the stereo data decreases with higher

distance.

In Table 2, the mean, the minimum and the max-

imum error for each parameter conﬁguration of sce-

nario 2 to scenario 4 are listed. The mean errors

show values ranging from 74 mm to 87 mm, which

is accurate enough for our application, because the

Activity Recognition based on High-Level Reasoning - An Experimental Study Evaluating Proximity to Objects and Pose Information

419

500 1000 1500 2000 2500 3000 3500

100

110

120

130

140

Distance from sensor in mm

Error in mm

Exposure time: 10 ms

Exposure time: 40 ms

Slow movement

Fast movement

Parallel to optical axis

Perpendicular to optical axis

Figure 3: Scenario 1. Inﬂuence of the distance on the

mean error for the deﬁned intervals. For distances higher

than 1000 mm, the error between the position measured by

the stereo camera and the reference position increases from

about 60 mm to a range from 100 mm to 130 mm.

distances between sink, shower and toilet are several

times higher.

Table 2: Determined mean, minimum and maximum errors

for scenario 2 to scenario 4. All numbers in mm.

Scenario Conﬁguration mean min. max.

Slow 74 2 339

Fast 83 2 654

Parallel 87 1 416

Perpendicular 76 1 842

Exp. 40 ms 83 1 396

Exp. 10 ms 86 2 398

In the following, the histograms are presented and

discussed for each scenario. In Figure 4, the results

for the two different moving speeds are presented.

Figure 4 and Table 2 show that the person detection

algorithm is more accurate for slower movements.

Figure 5 shows the results for the two different

moving directions.

In this experiment, we expected that the parallel

movement shows better results than the perpendicular

movement. This assumption is based on the fact, that

one stereo sensor just provides the 3-D points of the

part of a person that is visible to the sensor, i. e. half

of the person’s surface, either viewed from the side or

from the front or back respectively.

When the person is viewed from the side, the

points belonging to the arm and shoulder might cause

the center of these points to be located near the

shoulder instead near the actual body center (head).

When the person is viewed from the front or back re-

spectively, the center of the visible surface points is

more likely to be located near the actual body cen-

ter. Therefore, we expected that the center is shifted

< 100 100−200 200−300 300−400 400−500 > 500

Error in mm

Relative number in %

Slow movement

Fast movement

Figure 4: Histogram for scenario 2, moving speed.

< 100 100−200 200−300 300−400 400−500 > 500

Error in mm

Relative number in %

parallel to optical axis

perpendicular to optical axis

Figure 5: Histogram for scenario 3, moving direction.

to one shoulder when the person moves perpendicular

(viewed from the side), while the shift to the person’s

back or front is expected to be smaller when the per-

son walks parallel to the optical axis (viewed from

the front). In contrast to the expectation, the results of

the perpendicular movement direction are even better

than those of the parallel direction according to Fig-

ure 5.

Figure 6 shows the results for the two different

exposure times that represent different lighting con-

ditions in an AAL environment (bright scene vs. dark

scene).

In this test set-up, the person detection algorithm

performs slightly better in the brighter scene with the

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

420

< 100 100−200 200−300 300−400 400−500 > 500

Error in mm

Relative number in %

Exposure time: 10 ms

Exposure time: 40 ms

Figure 6: Histogram for scenario 4 (exposure time).

exposure time of 40 ms.

Generally, it can be observed that a conﬁguration

change has only a slight effect on the results. The his-

tograms show only a small difference for the param-

eter changes. Furthermore, Figure 3 shows that the

plotted lines are close to each other, which demon-

strates that the error is similarly high for the different

scenarios. Consequently, the experiments proved that

the person detection algorithm is rather robust against

these changes. During the experiments, systematic er-

rors are occurring. We assume that the major inaccu-

racies in the measurements originate from the cali-

bration procedures, i. e., intrinsic and stereo calibra-

tion as well as during the extrinsic calibration of the

stereo and the reference system. Even a small mis-

alignment of these coordinate systems will result in

a measurement inaccuracy that linearly increases for

higher distances. Random errors may occur when the

measurement system fails to detect the rigid body cor-

rectly because of occlusions or reﬂections.

5.2 Reasoning Algorithm

The following three ﬁgures illustrate the results for

each recorded sequence. The green-shaded bars rep-

resent the real, i. e. the manually labelled, activity.

The activity that was determined by the system is

marked with an orange line. The ordinate shows the

time in seconds, whereas the abscissa shows the ac-

tivity.

The graphs show that the reasoning algorithm

shows results of high quality even for the small bath

room in our testing ﬂat. During all sequences, only

few minor peaks with a false detection occur. Fur-

thermore, there is a small delay of only a few sec-

Time in seconds

0 20 40 60 80 100 120 140 160

Activity

Other

Using toilet

Washing hands

Taking a shower

Real Pose

Determined Pose

Figure 7: Test person 1: Comparison between labelled

(green) and real (orange) activity.

Time in seconds

0 20 40 60 80 100 120 140 160

Activity

Other

Using toilet

Washing hands

Taking a shower

Real Pose

Determined Pose

Figure 8: Test person 2: Comparison between labelled

(green) and real (orange) activity.

Time in seconds

0 20 40 60 80 100 120 140 160

Activity

Other

Using toilet

Washing hands

Taking a shower

Real Pose

Determined Pose

Figure 9: Test person 3: Comparison between labelled

(green) and real (orange) activity.

onds visible. The delay is caused by a Kalman ﬁlter

that is used as a tracker in the person detection al-

gorithm. Additionally, the reasoning algorithm is de-

signed as a low-pass ﬁlter in order to ﬁlter very small

peaks. In view of AAL applications with activities

Activity Recognition based on High-Level Reasoning - An Experimental Study Evaluating Proximity to Objects and Pose Information

421

spanning over a time of several minutes, neither the

peaks nor the delay have an effect on the function-

ality. To sum up the results, the reasoning algorithm

was able to detect all the activities that were part of the

reconstructed morning scene, except only a few small

peaks with a false detection. Thus far, no method can

be found in literature that addresses activity recogni-

tion related to hygiene aspects in AAL environments

in a comparable way.

6 CONCLUSIONS AND FUTURE

WORK

In this study, an algorithm for reasoning about ADLs

has been presented, which evaluates the proximity of

relevant objects as well as the person’s pose. In order

to determine whether the chosen person detection al-

gorithm is theoretically accurate enough – even when

objects are close to each other – the accuracy of the

algorithm has been analysed with respect to different

parameters relevant in AAL scenarios.

As a result, the algorithm has proved sufﬁcient

quality. It can consequently be stated that this algo-

rithm is appropriate for our AAL application, where

relevant objects are close to each other. Moreover,

the accuracy analysis has been designed in an univer-

sal fashion, so that other person detection algorithms

can be analysed under similar conditions, which fa-

cilitates comparisons. The experiments demonstrate

that the algorithm is accurate with regard to changing

conditions that prevail in AAL environments.

The evaluation of the reasoning algorithm in the

testing ﬂat demonstrated that activities normally per-

formed in front of a sink, such as ”washing hands,

combing, teeth brushing, etc.”, ”showering” and ”us-

ing the toilet” could be accurately determined. The

tests were conducted in the comparatively small bath-

room of our testing ﬂat, so that it can be assumed

that our approach would also show good results in

larger rooms. We plan to extend the algorithm to more

objects in other rooms, in order to recognise further

ADLs, such as ”preparing a meal”, ”washing up” or

”cooking”. In addition, we intend to conduct more

tests with probands in our testing ﬂat and to integrate

the designed system in real living environments. At

this point, we will continue working together with lo-

cal housing associations and care facilities.

In summary, technical support systems could con-

tribute to a higher quality of care. By giving advice,

sending reminding messages to patients and provid-

ing care-related information to caring personnel, these

modern developments could be beneﬁcial to patients,

caring personnel and relatives alike.

ACKNOWLEDGEMENTS

This project is funded by the European Social Fund

(ESF).

REFERENCES

Dayangac, E., Wiede, C., Richter, J., and Hirtz, G. (2015).

Robust Head-shoulder Detection using Deformable

Part-based Models. In 10th International Confer-

ence on Computer Vision Theory and Applications

(VISAPP-2015), pages 236–243.

Harville, M. and Li, D. (2004). Fast, integrated person

tracking and activity recognition with plan-view tem-

plates from a single stereo camera. In Computer Vi-

sion and Pattern Recognition, 2004. CVPR 2004. Pro-

ceedings of the 2004 IEEE Computer Society Confer-

ence on, volume 2, pages II–398. IEEE.

Pirsiavash, H. and Ramanan, D. (2012). Detecting activities

of daily living in ﬁrst-person camera views. In Com-

puter Vision and Pattern Recognition (CVPR), 2012

IEEE Conference on, pages 2847–2854. IEEE.

Richter, J., Christian, W., and Hirtz, G. (2015). Mobility

Assessment of Demented People Using Pose Estima-

tion and Movement Detection. In ICPRAM Lisbon,

Fourth International Conference on Pattern Recogni-

tion Applications and Methods, pages 22–29.

Richter, J., Findeisen, M., and Hirtz, G. (2014). Assess-

ment and Care System Based on People Detection

for Elderly Suffering From Dementia. In Consumer

Electronics Berlin (ICCE-Berlin), 2014. ICCE Berlin

2014. IEEE Fourth International Conference on Con-

sumer Electronics, pages 59–63. IEEE.

Scanaill, C. N., Carew, S., Barralon, P., Noury, N., Lyons,

D., and Lyons, G. M. (2006). A Review of Ap-

proaches to Mobility Telemonitoring of the Elderly in

Their Living Environment. Annals of Biomedical En-

gineering, 34(4):547–563.

Steen, E.-E., Frenken, T., Frenken, M., and Hein, A. (2013).

Functional Assessment in Elderlies Homes: Early Re-

sults from a Field Trial. Lebensqualit

at im Wandel von

Demograﬁe und Technik.

Yous, S., Laga, H., Chihara, K., et al. (2008). People de-

tection and tracking with world-z map from a single

stereo camera. In The Eighth International Workshop

on Visual Surveillance-VS2008.

Zivkovic, Z. (2004). Improved adaptive Gaussian mixture

model for background subtraction. In Pattern Recog-

nition, 2004. ICPR 2004. Proceedings of the 17th In-

ternational Conference on Pattern Recognition, vol-

ume 2, pages 28–31. IEEE.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

422