How Long Are Various Types of Daily Activities? Statistical Analysis of a

Multimodal Wearable Sensor-based Human Activity Dataset

Hui Liu

and Tanja Schultz

Cognitive Systems Lab, University of Bremen, Bremen, Germany

Keywords:

Human Activity, Human Activity Recogntion, Activities of Daily Living, Wearable Sensing, Internal Sensing,

Biosignals, CSL-SHARE, Statistical Analysis.

Abstract:

Human activity research in the ﬁeld of informatics, such as activity segmentation, modeling, and recognition,

is moving in an increasingly interpretable direction with the introduction of sports and kinematics knowledge.

Many related research topics face a question: How long is the typical duration of the activities needed to be

modeled? Several public human activity datasets do not strictly limit single motions’ repetition times, such

as gait cycle numbers, in recording sessions, so they are not statistically signiﬁcant concerning activity du-

ration. Standing on the rigorous acquisition protocol design and well-segmented data corpus of the recently

released multimodal wearable sensor-based human activity dataset CSL-SHARE, this paper analyzes the du-

ration statistics and distribution of 22 basic single motions of daily activities and sports, hoping to provide

research references for human activity studies. We discovered that (1) the duration of each studied human

daily activity or simple sports activity reﬂects interpersonal similarities and naturally obeys a normal distribu-

tion; (2) one single motion (such as jumping and sitting down) or one cycle in the activities of cyclical motions

(such as one gait cycle in walking) has an average duration in the interval from about 1 second to 2 seconds.

1 INTRODUCTION

In today’s highly automated society, human activities

are being studied more and more widely in the ﬁeld

of Artiﬁcial Intelligence (AI) to facilitate human life,

such as in medical care (Ejupi et al., 2016), interac-

tive interfaces (Ancans et al., 2017), and multime-

dia entertainment (Jung and Cha, 2010) (Zok, 2014).

Generally speaking, human activity can refer to all

behaviors related to human beings, such as brain ac-

tivity, which do not need to produce any movement.

However, when we use the concept of “human ac-

tivity” in the research category of Machine Learning

(ML), we generally mean the connotation of kinemat-

ics, for which the concept “motion” matches more

closely. More speciﬁcally, a human activity refers

to the movement(s) of one or several parts of the

person’s body, either atomic or composed of many

primitive actions performed in some sequential or-

der (Beddiar et al., 2020). Therefore, human activ-

ity has a broad denotation: It can refer to a single

human motion in a narrow sense, such as jumping,

walking, and most gestures, or a human motion se-

https://orcid.org/0000-0002-6850-9570

https://orcid.org/0000-0002-9809-7028

quence of concurrent, coupled, and sequential mo-

tions in a broad sense, like cutting a cake, as described

in (Gehrig, 2015). Moreover, in most public human

activity datasets, some postures which do not produce

substantial movement, such as standing and sitting,

are also categorized as the scope of (static) human

activity, since these activities can also be recognized

separately, given suitable equipment.

Many ML research topics related to human activ-

ities, such as segmentation, modeling, and recogni-

tion, face a question: How long is the duration of

the activity that needs to be modeled? For ofﬂine

activity modeling, the knowledge of activity duration

can help estimate the model parameters at the outset,

such as layer numbers in the Neural Network (NN) or

state numbers in the Hidden Markov Model (HMM).

For online research like real-time or streaming-based

Human Activity Recognition (HAR), the duration of

the activities to be recognized will help seek a bet-

ter trade-off between window length, window over-

lap, and performance delay (Liu and Schultz, 2019).

In sports science, activity duration is usually a

common measurement task, but there are few re-

search works where these measurement results are ad-

equately used in ML research. In kinematics, gait

680

Liu, H. and Schultz, T.

How Long Are Various Types of Daily Activities? Statistical Analysis of a Multimodal Wearable Sensor-based Human Activity Dataset.

DOI: 10.5220/0010896400003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 680-688

ISBN: 978-989-758-552-4; ISSN: 2184-4305

analysis provides a good model hypothesis (Mezghani

et al., 2013) (Whittle, 1996) (Whittle, 2014) (Arous

et al., 2018), but there are few related statistical ref-

erences based on signalized kinematic data, except

that (Lanshammar, 1987) estimated gait cycle du-

ration and stride length from one-marker kinematic

data. Therefore, it is essential to conduct a statisti-

cal activity duration analysis from the perspective of

human activity data mining.

Some human daily activities, such as sitting and

vacuuming, can last for half an hour or only a few sec-

onds. Hence, among all deﬁnitions and divisions of

human activities, two categories, in which the activ-

ity duration varies from person to person, from task to

task, and from dataset to dataset, are not statistically

informative in terms of duration:

• Postures, which can maintain a steady body state

for any duration, such as standing, sitting, lying,

squatting, among others;

• Activities of sequential motions, such as cooking,

vacuuming, watching TV, among others.

This paper focuses on analyzing the duration

statistics and distribution of single motions. No mat-

ter the motion contains only one movement, such as

one-leg jumping, two-leg jumping, standing up, sit-

ting down, and turning left/right, or several contin-

uous looping cycles, such as walking, running, go-

ing upstairs/downstairs, and lateral shufﬂing, its du-

ration has the statistical value of interpersonal simi-

larity, which not only provides a powerful reference

for modeling but also may facilitate processing more

complex motion sequences.

2 DATASET

Except for some human activity datasets that involve

relatively small numbers of deﬁned activities, such as

RealWorld (Sztyler and Stuckenschmidt, 2016), EN-

ABL3S (Hu et al., 2018), and Gait Analysis Data Base

(Loose et al., 2020), or only activities of particu-

lar body parts, such as mmGest for gesture (Georgi

et al., 2015), CSL hdemg for ﬁnger (Amma et al.,

2015), and Upper-body movements (Santos et al.,

2020), most of the existing public human activity

datasets, such as OPPORTUNITY (Roggen et al.,

2010), PAMAP2, (Reiss and Stricker, 2012), Daily

Log (Sztyler et al., 2016), and FORTH-TRACE (Kara-

giannaki et al., 2016), cannot be directly applied to the

statistical analysis of activity duration because of the

following two reasons:

• Unrestricted acquisition protocol designs. For ex-

ample, the activity “walking” itself can last for

any duration, but each gait cycle of a healthy adult

can be statistically analyzed; however, there are

few datasets stipulating each “walking” segment

with a ﬁxed number of gait cycles strictly;

• Distinct activity segmentation methods. For ex-

ample, the UniMiB SHAR dataset (Micucci et al.,

2017) implements a simple way of ﬁnding the

magnitude peak of the acceleration signals to seg-

ment 17 classes of Activities of Daily Living

(ADLs) and falls. Whether it is walking, jump-

ing, or falling forward, each activity segment is

precisely 3 seconds. This kind of segmentation is

easy and efﬁcient with almost no manual labor or

machine learning study but can hardly be applied

for accurate human activity duration analysis.

The data support we use is a multimodal wear-

able sensor-based human activity dataset called CSL-

SHARE (Liu et al., 2021a), whose quality and appli-

cability have been extensively veriﬁed in many re-

search ﬁelds of human activities, such as HAR re-

search pipeline (Liu et al., 2022), feature extraction

(Barandas et al., 2020), feature space reduction (Hart-

mann et al., 2021) (Hartmann et al., 2022), automatic

segmentation, human activity modeling and recogni-

tion (Liu et al., 2021b), among others. A knee ban-

dage was used as a wearable sensor carrier, making

the dataset distinctive and more kinematically signif-

icant. The 19-channel dataset was recorded from 9

biomechanical and bioelectrical sensors, including 2

triaxial accelerometers, 2 triaxial gyroscopes, 4 EMG

sensors, 1 biaxial electrogoniometer, and 1 airborne

microphone with sampling rates up to 1,000 Hz. By

applying the in-house implemented software Activity

Signal Kit (ASK) (Liu and Schultz, 2018) for data

acquisition, segmentation, and annotation, the CSL-

SHARE dataset covers 22 types of ADLs and sports

from 20 subjects, 5 female and 15 male, aged between

23 and 43 (30.5 ± 5.8), in a total time of 11.52 hours,

of which 6.05 hours are segmented and annotated.

Unlike many human activity datasets listed above,

the CSL-SHARE dataset adopts strictly deﬁned ac-

quisition protocols and a semi-automatic segmen-

tation mechanism called “protocol-for-pushbutton”,

enabling efﬁcient and accurate statistical analysis

of single motions. In addition to strictly stipulat-

ing “three gait cycles” and “left-foot-ﬁrst” for gait-

based activities such as walking, running, going up-

stairs/downstairs, and left/right lateral shufﬂing, it

is worth mentioning that the CSL-SHARE dataset

also distinguishes some activities with left-foot-ﬁrst

or right-foot-ﬁrst, providing materials for similarity

analysis of activity duration.

How Long Are Various Types of Daily Activities? Statistical Analysis of a Multimodal Wearable Sensor-based Human Activity Dataset

681

Table 1: Statistics of the single motion segment duration in the CSL-SHARE dataset. The minimum, maximum, mean, and

standard deviation (std.) values are in seconds.

Activity Minimum Maximum Mean±std. Number of segments

jump-one-leg 0.830 2.949 1.69 ± 0.33 379

jump-two-leg 0.869 3.389 1.95 ± 0.39 380

walk (one gait cycle) 1.046 1.863 1.42 ± 0.15 400

walk-curve-left 90

◦

(one gait cycle) 0.966 2.150 1.45 ± 0.19 398

walk-curve-right 90

◦

(one gait cycle) 1.076 2.063 1.48 ± 0.17 393

walk-upstairs (one gait cycle) 1.263 2.243 1.59 ± 0.15 365

walk-downstairs (one gait cycle) 1.023 1.973 1.44 ± 0.17 364

spin-left-left-ﬁrst 0.959 3.069 1.67 ± 0.30 380

spin-left-right-ﬁrst 0.969 2.609 1.83 ± 0.29 420

spin-right-left-ﬁrst 0.800 2.619 1.86 ± 0.24 401

spin-right-right-ﬁrst 1.169 2.719 1.71 ± 0.22 400

run (one gait cycle) 0.773 1.373 1.05 ± 0.11 400

shufﬂe-left (one gait cycle) 0.580 1.290 0.96 ± 0.10 380

shufﬂe-right (one gait cycle) 0.696 1.386 0.97 ± 0.11 374

V-cut-left-left-ﬁrst 0.809 3.039 1.81 ± 0.33 399

V-cut-left-right-ﬁrst 1.019 2.709 1.88 ± 0.29 378

V-cut-right-left-ﬁrst 0.840 2.759 1.80 ± 0.34 400

V-cut-right-right-ﬁrst 1.209 2.649 1.84 ± 0.28 378

sit-to-stand 1.049 2.719 1.81 ± 0.32 389

stand-to-sit 1.129 3.729 1.92 ± 0.35 389

sit 0.819 8.019 1.66 ± 0.58 389

stand 0.809 6.959 1.64 ± 0.51 405

3 ACTIVITY DURATION

ANALYSIS

Table 1 gives the number of activity segments and the

minimal/maximal/mean duration of the 22 activities

in the CSL-SHARE dataset. It should be noted that,

unlike the statistics of the table given in (Liu et al.,

2021a), for the eight activities involving three com-

plete gait cycles, we only list the statistics of one gait

cycle to create referenceability for other research.

Judging from each activity segment’s number of

occurrences, we can ﬁnd that this dataset is well-

balanced for each involved activity, reﬂecting strict

protocol design and execution. Each activity was

planned to be performed 20 times by each participant

according to the protocols. The activity occurrence

discrepancy in Table 1 is mainly due to eliminating

the misoperation during users’ execution of the semi-

automatic segmentation mechanism.

Since the pushbutton for segmentation and anno-

tation may be pressed/released earlier or later during

the acquisition process, millisecond-level operation-

related slight duration deviations make the minimum

and maximum values’ general statistical reference of

minor signiﬁcance; however, they are still meaning-

ful when using the CSL-SHARE dataset for model-

ing research. Due to the big data effect of multiple

participants and multiple activity execution times, the

impact of individual operational discrepancy is com-

pensated to a high degree on the average/standard de-

viation values in Table 1. Furthermore, it is notewor-

thy that the video-based manual post veriﬁcation after

each data collection process corrected obvious dura-

tion outliers.

From Table 1 we can deduce a statistical hypoth-

esis: One single motion (such as jumping and sit-

ting down) or one cycle in the cyclical single motions

(such as one gait cycle in walking) has an average du-

ration in the interval from about 1 second to 2 sec-

onds, which can help determine a priori some param-

eters for data segmentation, feature extraction, activ-

ity modeling, model training, and recognition (Hart-

mann et al., 2020) (Liu, 2021). It must be stressed that

the dataset was recorded from only healthy young to

middle-aged adults. In fact, except for unique appli-

cation scenarios, ML research often starts at this age

group. Moreover, for biomedical engineering, data

of healthy individuals is usually the ﬁrst material for

establishing an applicable model, which also creates

reliable references for the study of pathological situa-

tions.

HEALTHINF 2022 - 15th International Conference on Health Informatics

682

The following subsections will analyze the dura-

tion statistics of each activity group and their distribu-

tion.

Figure 1: Duration histograms of the human activi-

ties jump-one-leg and jump-two-leg in the CSL-SHARE

dataset. The area under the curve equals the total number of

segment occurrences within 100-millisecond intervals.

3.1 Jumping Activities

The acquisition protocols for single-leg and two-leg

jumping activities in the CSL-SHARE dataset are de-

scribed as “squat, then jump upwards using the ban-

daged right leg/both legs, land in” (Liu et al., 2021a).

Table 1 and Figure 1 demonstrate that the aver-

age duration of the two jumping activities is within

the interval of 1.6 – 2.0 seconds, and the duration of

a single-leg jump is about 86% shorter than that of

a two-leg jump. It is more challenging to keep bal-

ance during the single-leg squat, so subjects gener-

ally shortened the time and amplitude of the single-

leg squat and tended to jump as soon as possible.

Therefore, the relatively reduced muscular power in

the single-leg jump also shortens the body’s stay-in-

the-air (ﬂy) time, compared to a well-prepared two-

leg jump. This phenomenon is also witnessed by their

half-second maximum duration difference, while the

minimum values are close. Participants tended to be

more prepared for the squat in two-leg jump.

Jumping activities are certainly not only restricted

to the direction of upwards — it can be forwards-

upwards, or even leftwards/rightwards. Physically

speaking, regardless of what kind of jump happens,

as long as it happens daily and generally without a

particular purpose like a header shot, there should be

no apparent difference in duration statistics.

Figure 2: Duration histograms of one gait cycle in the

cyclical motions walk, walk-curve-left, walk-curve-right,

walk-upstairs, and walk-downstairs in the CSL-SHARE

dataset. The area under the curve equals the total number of

segment occurrences within 100-millisecond intervals.

How Long Are Various Types of Daily Activities? Statistical Analysis of a Multimodal Wearable Sensor-based Human Activity Dataset

683

3.2 Activities of Gait-based Cyclical

Motions at Waling Speed

The acquisition protocols for the ﬁve gait-based cycli-

cal motions in the CSL-SHARE dataset are described

as follows (Liu et al., 2021a):

• Walk: walk forward with three gait cycles, left

foot starts, i.e., left-right-left-right-left-right;

• Walk-curve-left/right: turn left/right 90

◦

with

three gait cycles at walking speed, left foot starts;

• Walk-upstairs/downstairs: go up/down six stairs

with three gait cycles, left foot starts.

As mentioned above, in order to establish a univer-

sal reference, we only describe one-gait-cycle statis-

tics in Table 1 and Figure 2.

Regardless of the direction, the duration of a

gait cycle at daily walking speed is about 1.4 to 1.6

seconds. Obviously, when turning left/right or go-

ing upstairs/downstairs during walking, the average

gait cycle duration is slightly longer than the normal

straightforward walking, among which walking up-

stairs leads a duration of about 100 – 200 millisec-

onds longer on average. It can also be observed that

although walking downstairs saves 10% of the time

than upstairs, it is not faster than walking straightfor-

ward. In real life, walking downstairs is considered

a fast movement than other types of walking, but in

the laboratory data acquisition sessions, participants

tended to go downstairs at a relatively normal speed

in a relaxed environment. Another thing to remind is

that the height of the stairs will also affect the activity

duration. The stairs used in the CSL-SHARE dataset

are the regular building stairs of standard height.

3.3 Single-gait Activities at Walking

Speed

The “spin-left” and “spin-right” activities in the CSL-

SHARE dataset can be described as the “Left face!”

or “Right face!” action in the army (but in daily situa-

tions, not so stressful as in military training). The ac-

quisition protocols are designed as “turn left/right 90

◦

in one step, left/right foot starts” (Liu et al., 2021a).

“Spin-left” is divided into “spin-left-left-ﬁrst” and

“spin-left-right-ﬁrst,” denoting which foot should be

moved ﬁrst. Similarly, “spin-right” is also divided

into two activities in regard to the ﬁrst-moved foot.

The reason for the subdivision is that these activities

only involve one gait cycle, and the data acquisition

only uses the sensors placed on the right-leg-worn

bandage. Therefore, the “left-foot-ﬁrst” and “right-

foot- ﬁrst” of these activities will lead to very differ-

Figure 3: Duration histograms of the single-gait human ac-

tivities spin-left-left-ﬁrst, spin-left-right-ﬁrst, spin-right-

left-ﬁrst, and spin-right-right-ﬁrst in the CSL-SHARE

dataset. The area under the curve equals the total number

of segment occurrences within 100-millisecond intervals.

ent signal patterns. On the contrary, activities involv-

ing multiple gait cycles (see Sections 3.2 and 3.4) are

not subdivided according to the ﬁrst-moving foot.

As Table 1 and Figure 3 exhibit, the average dura-

tion of 90

◦

single-gait turns is larger than a single gait

cycle in walking activities due to the large turning an-

gles. The average duration is about 1.6 – 1.9 seconds.

It takes 9% more time to start with the right foot in a

left turn or start with the left foot in a right turn than to

start with the same side foot in the turning direction.

HEALTHINF 2022 - 15th International Conference on Health Informatics

684

Figure 4: Duration histograms of one gait cycle in the cycli-

cal motions run, shufﬂe-left, and shufﬂe-right in the CSL-

SHARE dataset. The area under the curve equals the total

number of segment occurrences within 100-millisecond in-

tervals.

3.4 Activities of Gait-based Cyclical

Motions at Fast Speed

The acquisition protocol of “run” is basically the

same as that of “walk,” except the fast speed, while

the protocols of “shufﬂe-left/right” demand the sub-

ject to “move leftward/rightward with three lateral

gaits cycles, left/right foot starts, the other foot fol-

lows” (Liu et al., 2021a).

One gait cycle of the sports-related lateral shuf-

ﬂing, as Table 1 and Figure 4 display, is the fastest

motion in the entire CSL-SHARE dataset, followed by

running. Their duration is around 1 second. Limited

to laboratory conditions, running was actually per-

formed at jogging speed instead of reaching full speed

in sports.

Roughly speaking, the duration of a single gait at a

fast-paced speed is about 30% shorter than at a walk-

ing speed. It can be highlighted that these three short-

est motions’ statistics present the least standard devi-

ations from subject to subject in the whole dataset.

3.5 Single-gait Activities at Fast Speed

“V-cut” in the CSL-SHARE dataset refers to the

single-gait motion of direction changing during run-

ning.

Figure 5: Duration histograms of the single-gait human

activities v-cut-left-left-ﬁrst, v-cut-left-right-ﬁrst, v-cut-

right-left-ﬁrst, and v-cut-right-right-ﬁrst in the CSL-

SHARE dataset. The area under the curve equals the total

number of segment occurrences within 100-millisecond in-

tervals.

How Long Are Various Types of Daily Activities? Statistical Analysis of a Multimodal Wearable Sensor-based Human Activity Dataset

685

The acquisition protocols of “V-cut” are stipulated

as follows: “turn 30

◦

left/right forward in one step

at jogging speed, left/right foot starts” (Liu et al.,

2021a). Similar to the “spin” activities, both “V-cut-

left” and “V-cut-right” are divided into two activities

regarding the ﬁrst-moved foot, separately, due to the

same reason as explained in Section 3.3.

The average duration of all four “V-cut” activities

falls within 1.8 seconds to 1.9 seconds, as Table 1 and

Figure 5 indicate. Compared to the fact that there is a

signiﬁcant duration difference of which foot is moved

ﬁrst in the “spin” activities, for both “V-cut-left” and

“V-cut-right”, which foot is moved ﬁrst affect only

slightly the activity duration, for which two reasons

can explain: Firstly, the “V-cut” activities themselves

require a large step with intense muscular prepara-

tion, where which foot to step ﬁrst has no signiﬁcant

inﬂuence; Secondly, it is stated in the protocol that

by “V-cut”, subjects only need to rotate 30

◦

, which

is one-third of the 90

◦

in spin activities (this is also

in line with common sense — it’s easy to turn 90

◦

while walking, but directional changing at fast speed

requires a larger motion radius).

Interestingly, whether turning 30 degrees to the

left or the right, starting with the right foot always

causes a little longer duration in average.

3.6 Activities of Transition between

Standing and Sitting

Figure 6 illustrates that the “stand-to-sit” activity, i.e.,

sitting down, has a similar duration distribution to

“sit-to-stand”, i.e., standing up. However, according

to Table 1 and Figure 6, it is noticeable that averagely,

sitting down is about 100 milliseconds longer than

standing up and has more outlier samples with long

duration, which is consistent with real-life situations:

The knee ﬂexion is more difﬁcult to act than the knee

extension, and a certain sense of organ self-protection

often accompanies knee bent.

3.7 Postures: Standing and Sitting

Activities

Like many public human activity datasets (see Section

2), CSL-SHARE also includes standing and sitting ac-

tivities, which should be classiﬁed as postures. Never-

theless, as mentioned in Section 1, these two activities

are considered no signiﬁcant interindividual statisti-

cal reference, for which some clues can be glimpsed

through the long duration outliers in Figure 7. Due to

the outlier maximum duration values, these two his-

tograms use a 200-millisecond interval, different from

Figure 6: Duration histograms of the human activities sit-

to-stand and stand-to-sit in the CSL-SHARE dataset. The

area under the curve equals the total number of segment

occurrences within 100-millisecond intervals.

Figure 7: Duration histograms of the segmented postures

stand and sit in the CSL-SHARE dataset. The area under

the curve equals the total number of segment occurrences

within 200-millisecond intervals.

the 100-millisecond interval of other activities’ dura-

tion histograms, in order to display the horizontal axis

more clearly.

An arresting point can still be discovered. Even

if the organizer did not specify the duration of each

standing/sitting acquisition segment, and the partic-

ipants did not observe each other, the vast majority

HEALTHINF 2022 - 15th International Conference on Health Informatics

686

of participants still performed each standing and sit-

ting at about 1 to 2 seconds using the pushbutton, and

the overall duration statistics of these two activities

are approximately normally distributed. Such phe-

nomenons may involve a variety of topics such as

behavioral science and natural psychological rhythm,

which will not be expanded due to the different re-

search ﬁelds.

4 CONCLUSIONS

Relying on the rigorous acquisition protocol design

and execution, as well as the well-segmented data

corpus of the recently released multimodal wearable

sensor-based human activity dataset CSL-SHARE,

this paper analyzes the duration statistics and distri-

bution of 22 basic single motions of daily activities

and sports, providing research references for human

activity studies, such as segmentation, feature extrac-

tion, modeling, and recognition.

Through the big data statistical analysis of each

activity’s duration, we discovered that one single-

motion activity or one cycle in the activities of cycli-

cal motions has an average duration in the interval

from about 1 second to 2 seconds.

Furthermore, the duration distribution histograms

of each studied human daily activity or simple sports

activity evince interindividual similarities and natu-

rally obey a normal distribution. Even the two pos-

tures, standing and sitting, for which participants ar-

bitrarily decided each segment’s length, also conform

to this observation unpredictably.

As a classic case of applying activity duration

statistics in ML, (Liu and Schultz, 2019) used the pre-

vious dataset of CSL-SHARE with the same equip-

ment and investigated the transition from the ofﬂine

HAR modeling research to a real-time HAR system.

The activity duration was utilized as one of the ref-

erences to ﬁnd the optimal balance between the on-

line decoding window length, the window overlap

length, and the recognition delay, endowing the real-

time demonstration with a satisfactory performance

and user experience.

We have noticed that different types of falling ac-

tivities also show interindividual similarity in terms

of duration, which is of great signiﬁcance for human

activity research based on internal sensing and exter-

nal sensing, such as adopting HAR modeling for fall

detection and recognition (Xue and Liu, 2021). Dura-

tion analysis of typical falling activities will be a valu-

able topic to explore in the future, given appropriate

and adequate research materials.

REFERENCES

Amma, C., Krings, T., B

oer, J., and Schultz, T. (2015).

Advancing muscle-computer interfaces with high-

density electromyography. In Proceedings of the 33rd

Annual ACM Conference on Human Factors in Com-

puting Systems, pages 929–938, New York, NY, USA.

ACM.

Ancans, A., Rozentals, A., Nesenbergs, K., and Greitans,

M. (2017). Inertial sensors and muscle electrical sig-

nals in human-computer interaction. In ICTA 2017

- 6th International Conference on Information and

Communication Technology and Accessibility, pages

1–6. IEEE.

Arous, M. A. B., Dunbar, M., Arfaoui, S., Mitiche,

A., Ouakrim, Y., Fuentes, A., Richardson, G., and

Mezghani, N. (2018). Knee kinematics feature selec-

tion for surgical and nonsurgical arthroplasty candi-

date characterization. In Proceedings of the 11th In-

ternational Joint Conference on Biomedical Engineer-

ing Systems and Technologies - Volume 3: BIOSIG-

NALS, pages 176–181.

Barandas, M., Folgado, D., Fernandes, L., Santos, S.,

Abreu, M., Bota, P., Liu, H., Schultz, T., and Gam-

boa, H. (2020). TSFEL: Time series feature extraction

library. SoftwareX, 11:100456.

Beddiar, D. R., Nini, B., Sabokrou, M., and Hadid,

A. (2020). Vision-based human activity recogni-

tion: A survey. Multimedia Tools and Applications,

79(41):30509–30555.

Ejupi, A., Gschwind, Y. J., Valenzuela, T., Lord, S. R., and

Delbaere, K. (2016). A kinect and inertial sensor-

based system for the self-assessment of fall risk: A

home-based study in older people. HCI 2016 - 18th

International Conference on Human-Computer Inter-

action, 31(3-4):261–293.

Gehrig, D. (2015). Automatic Recognition of Concurrent

and Coupled Human Motion Sequences. PhD thesis,

Karlsruher Institut f

ur Technologie.

Georgi, M., Amma, C., and Schultz, T. (2015). Recognizing

hand and ﬁnger gestures with imu based motion and

emg based muscle activity sensing. In Proceedings of

the International Conference on Bio-inspired Systems

and Signal Processing - BIOSIGNALS, (BIOSTEC

2015), pages 99–108. INSTICC, SciTePress.

Hartmann, Y., Liu, H., Lahrberg, S., and Schultz, T. (2022).

Interpretable high-level features for human activity

recognition. In BIOSIGNALS 2022 — 15th Interna-

tional Conference on Bio-inspired Systems and Signal

Processing. INSTICC, SciTePress. forthcoming.

Hartmann, Y., Liu, H., and Schultz, T. (2020). Feature space

reduction for multimodal human activity recognition.

In Proceedings of the 13th International Joint Confer-

ence on Biomedical Engineering Systems and Tech-

nologies - Volume 4: BIOSIGNALS, pages 135–140.

INSTICC, SciTePress.

Hartmann, Y., Liu, H., and Schultz, T. (2021). Feature

space reduction for human activity recognition based

on multi-channel biosignals. In Proceedings of the

14th International Joint Conference on Biomedical

How Long Are Various Types of Daily Activities? Statistical Analysis of a Multimodal Wearable Sensor-based Human Activity Dataset

687

Engineering Systems and Technologies, pages 215–

222. INSTICC, SciTePress.

Hu, B., Rouse, E., and Hargrove, L. (2018). Benchmark

datasets for bilateral lower-limb neuromechanical sig-

nals from wearable sensors during unassisted locomo-

tion in able-bodied individuals. Frontiers in Robotics

and AI, 5:14.

Jung, Y. and Cha, B. (2010). Gesture recognition based on

motion inertial sensors for ubiquitous interactive game

contents. IETE Technical Review, 27(2):158–166.

Karagiannaki, K., Panousopoulou, A., and Tsakalides, P.

(2016). The forth-trace dataset for human activity

recognition of simple activities and postural transi-

tions using a body area network.

Lanshammar, H. (1987). Estimation of gait cycle dura-

tion and stride length from one marker kinematic data.

In Biomechanics: Basic and Applied Research, pages

129–134. Springer.

Liu, H. (2021). Biosignal Processing and Activity Model-

ing for Multimodal Human Activity Recognition. PhD

thesis, University of Bremen.

Liu, H., Hartmann, Y., and Schultz, T. (2021a). CSL-

SHARE: A multimodal wearable sensor-based human

activity dataset. Frontiers in Computer Science, 3:90.

Liu, H., Hartmann, Y., and Schultz, T. (2021b). Mo-

tion Units: Generalized sequence modeling of hu-

man activities for sensor-based activity recognition. In

EUSIPCO 2021 — 29th European Signal Processing

Conference, pages 1506–1510.

Liu, H., Hartmann, Y., and Schultz, T. (2022). A practical

wearable sensor-based human activity recognition re-

search pipeline. In BIODEVICES 2022 — 15th Inter-

national Conference on Biomedical Electronics and

Devices. INSTICC, SciTePress. forthcoming.

Liu, H. and Schultz, T. (2018). ASK: A framework for data

acquisition and activity recognition. In Proceedings of

the 11th International Joint Conference on Biomedi-

cal Engineering Systems and Technologies - Volume 3:

BIOSIGNALS, pages 262–268. INSTICC, SciTePress.

Liu, H. and Schultz, T. (2019). A wearable real-time hu-

man activity recognition system using biosensors in-

tegrated into a knee bandage. In Proceedings of the

12th International Joint Conference on Biomedical

Engineering Systems and Technologies - Volume 1:

BIODEVICES, pages 47–55. INSTICC, SciTePress.

Loose, H., Tetzlaff, L., and Bolmgren, J. L. (2020). A public

dataset of overground and treadmill walking in healthy

individuals captured by wearable imu and semg sen-

sors. In Biosignals 2020 - 13th International Confer-

ence on Bio-Inspired Systems and Signal Processing,

pages 164–171. INSTICC, SciTePress.

Mezghani, N., Fuentes, A., Gaudreault, N., Mitiche, A.,

Aissaoui, R., Hagmeister, N., and De Guise, J. A.

(2013). Identiﬁcation of knee frontal plane kinematic

patterns in normal gait by principal component anal-

ysis. Journal of Mechanics in Medicine and Biology,

13(03):1350026.

Micucci, D., Mobilio, M., and Napoletano, P. (2017).

UniMiB SHAR: A dataset for human activity recog-

nition using acceleration data from smartphones. Ap-

plied Sciences, 7(10):1101.

Reiss, A. and Stricker, D. (2012). Introducing a new bench-

marked dataset for activity monitoring. In ISWC 2012

- 16th International Symposium on Wearable Comput-

ers, pages 108–109. IEEE.

Roggen, D., Calatroni, A., Rossi, M., Holleczek, T., F

orster,

K., Tr

oster, G., Lukowicz, P., Bannach, D., Pirkl, G.,

Ferscha, A., et al. (2010). Collecting complex activ-

ity datasets in highly rich networked sensor environ-

ments. In INSS 2010 - 7th International Conference

on Networked Sensing Systems, pages 233–240. IEEE.

Santos, S., Folgado, D., and Gamboa, H. (2020). Upper-

body movements: Precise tracking of human motion

using inertial sensors.

Sztyler, T., Carmona, J., V

olker, J., and Stuckenschmidt,

H. (2016). Self-tracking reloaded: Applying process

mining to personalized health care from labeled sen-

sor data. 9930:160–180.

Sztyler, T. and Stuckenschmidt, H. (2016). On-body lo-

calization of wearable devices: An investigation of

position-aware activity recognition. In PerCom 2016

- 14th IEEE International Conference on Pervasive

Computing and Communications, pages 1–9. IEEE

Computer Society.

Whittle, M. W. (1996). Clinical gait analysis: A review.

Human movement science, 15(3):369–387.

Whittle, M. W. (2014). Gait analysis: An introduction.

Butterworth-Heinemann.

Xue, T. and Liu, H. (2021). Hidden Markov Model and its

application in human activity recognition and fall de-

tection: A review. In CSPS 2021 - 10th International

Conference on Communications, Signal Processing,

and Systems.

Zok, M. (2014). Inertial sensors are changing the games.

In ISISS 2014 - International Symposium on Inertial

Sensors and Systems, pages 1–3. IEEE.

HEALTHINF 2022 - 15th International Conference on Health Informatics

688