Feature Space Reduction for Human Activity Recognition

based on Multi-channel Biosignals

Yale Hartmann, Hui Liu and Tanja Schultz

Cognitive Systems Lab, University of Bremen, Germany

Keywords:

Human Activity Recognition, Biosignals, Multi-channel Signal Processing, Feature Space Reduction,

Stacking.

Abstract:

In this paper, we study the effect of Feature Space Reduction for the task of Human Activity Recognition

(HAR). For this purpose, we investigate a Linear Discriminant Analysis (LDA) trained with Hidden Markov

Models (HMMs) force-aligned targets. HAR is a typical application of machine learning, which includes ﬁnd-

ing a lower-dimensional representation of sequential data to address the curse of dimensionality. This paper

uses three datasets (CSL19, UniMiB, and CSL18), which contain data recordings from humans performing

more than 16 everyday activities. Data were recorded with wearable sensors integrated into two devices, a

knee bandage and a smartphone. First, early-fusion baselines are trained, utilizing an HMM-based approach

with Gaussian Mixture Models to model the emission probabilities. Then, recognizers with feature space

reduction based on stacking combined with an LDA are evaluated and compared against the baseline. Ex-

perimental results show that feature space reduction improves balanced accuracy by ten percentage points on

the UniMiB and seven points on the CSL18 datasets while remaining the same on the CSL19 dataset. The

best recognizers achieve 93.7 ± 1.4% (CSL19), 69.5 ± 8.1% (UniMiB), and 70.6 ± 6.0% (CSL18) balanced

accuracy in a leave-one-person-out cross-validation.

1 INTRODUCTION

The curse of dimensionality haunts all machine learn-

ing problems. Each feature dimension adds exponen-

tially to the data required to train a machine learn-

ing algorithm effectively, and various techniques have

been used and evaluated in a Human Activity Recog-

nition (HAR) context to address this explosion in the

required data. Feature selection using ﬁlter meth-

ods like minimum Redundancy Maximum Relevance

(mRMR), analysis of variance (ANOVA), and wrap-

per methods like forward selection, have been stud-

ied. They are known to successfully increase recog-

nition performance in many ﬁelds including HAR and

Automatic Speech Recognition (ASR) tasks (Bulling

et al., 2014) (Weiner and Schultz, 2018) (Capela

et al., 2015) (Suto et al., 2016). Feature space trans-

formation methods like Linear Discriminant Analy-

sis (LDA), Principal Component Analysis (PCA), and

Autoencoders (AE) have also been applied to HAR

with success (Almaslukh et al., 2018) (Mezghani

et al., 2013) (Li et al., 2018) (Hu and Zahorian, 2010).

However, most of these methods are designed for

single vector-based machine learning problems, not

sequential data, often found in HAR tasks. Com-

plex everyday activities, for example, are rarely deter-

mined properly by a snapshot of the human moving

around while performing this activity. Sitting down

and standing up look the same in a still image. Be-

side context, HAR tasks also require modeling and

transformation techniques adequate for sequences of

different lengths. Feature selection is possible with

a sequential classiﬁer in a wrapper method like for-

ward selection, and transformation can be done using

AE. Alternatively, transformation is also possible by

re-labeling the sequential data to a vector-based prob-

lem using Hidden Markov Model (HMM) force align-

ments and assigning each vector of the sequence to a

distinct state. These alignment-based reductions have

been applied in ASR tasks (Haeb-Umbach and Ney,

1992) (Siohan, 1995).

Our previous work adapted this alignment-based

technique for an HAR task and showed improvements

in recognition accuracy (Hartmann et al., 2020). This

work evaluates the reduction technique on two larger

datasets, details improvements, and compares the

achieved performances with the state of the art.

Hartmann, Y., Liu, H. and Schultz, T.

Feature Space Reduction for Human Activity Recognition based on Multi-channel Biosignals.

DOI: 10.5220/0010260802150222

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 215-222

ISBN: 978-989-758-490-9

 2021 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reserved

215

2 DATASETS

In this paper, three datasets, CSL18, CSL19, and

UniMiB, are used. They provide segmented and an-

notated sequential human activity data recorded with

wearable sensors or smartphones. The following sec-

tion provides a brief overview of the three datasets.

2.1 CSL18 and CSL19

The CSL18-18A-21S-4P (CSL18) in-house four-

person dataset recorded 20 biosignals and distin-

guishes 18 different activities. It contains 40mins of

semi-automatic annotated data recorded from sensors

attached to a knee bandage and was recorded at the

Cognitive Systems Lab (CSL) in a controlled labora-

tory environment.

The CSL19-22A-17S-20P (CSL19) in-house

dataset is a follow up to the CSL18 dataset. The

22-activity dataset contains six hours of segmented

and annotated data from 20 participants. In contrast

to CSL18, the two microphones and one goniometer

channel are dropped as they do not contain useful

information. The activities (listed in Table 1) are

comprised of everyday and sports activities chosen

for their relevance to gonarthrosis and personal

monitoring of knee straining behavior. The sensors

are placed on a knee bandage, as gonarthrosis patients

wear them for support and pain relief.

The sensors include four electromyographic sen-

sors (EMG) to capture muscle activity, as well as a

goniometer and two 3D-inertial measurement units

(IMU) consisting of a triaxial accelerometer and gy-

roscope to capture motion. 17 biosignals with EMG

sensors sampled at 1000Hz and IMUs and goniometer

sampled at 100Hz are captured.

We chose the biosignalsplux Research Kit

as a

recording device. One PLUX hub can process signals

from 8 channels (each up to 16 bits) simultaneously.

Therefore, three synchronized hubs are used during

the entire session. For more detail on the setup and

segmentation, please refer to (Liu and Schultz, 2018)

and (Liu and Schultz, 2019).

2.2 UniMiB-SHAR

The 17-activity UniMiB-SHAR (UniMiB) dataset has

been recorded at the University Milano Bicocca and

published as a benchmark dataset (Micucci et al.,

2017). The dataset focuses on everyday activities as

well as different types of falls listed in Table 2. It con-

tains nine different Activities of Daily Living (ADLs)

and eight different falls from 30 subjects. A triaxial

biosignalsplux.com/products/kits/researcher.html

Table 1: Overview of recorded activities in CSL19 dataset.

Durations given in seconds.

Target Min. Occ. Total

Walk 3.138 400 1702.076

Walk upstairs 3.788 365 1736.030

Walk downstairs 3.068 364 1568.212

Walk 90

◦

-curve left 2.898 398 1725.124

Walk 90

◦

-curve right 3.228 393 1749.744

Spin left left-ﬁrst 0.958 380 633.080

Spin left right-ﬁrst 0.968 420 767.409

Spin right left-ﬁrst 0.799 401 745.087

Spin right right-ﬁrst 1.168 400 685.610

V-Cut left left-ﬁrst 0.808 399 722.482

V-Cut left right-ﬁrst 1.018 378 709.474

V-Cut right left-ﬁrst 0.839 400 718.664

V-Cut right right-ﬁrst 1.208 378 695.446

Shufﬂe left 1.738 380 1097.810

Shufﬂe right 2.088 374 1089.572

Run 2.318 400 1260.279

Jump one leg 0.829 379 639.363

Jump two legs 0.868 380 739.360

Sit 0.818 389 646.282

Stand 0.808 405 663.870

Sit down 1.128 389 746.222

Stand up 1.048 389 705.052

Table 2: Overview of recorded activities in UniMiB dataset.

Durations given in seconds.

Target Dur. Occ. Total

FallingBack 3.0 526 1578.0

FallingBackSC 3.0 434 1302.0

FallingForw 3.0 529 1587.0

FallingLeft 3.0 534 1602.0

FallingRight 3.0 511 1533.0

FallingWithPS 3.0 484 1452.0

GoingDownS 3.0 1324 3972.0

GoingUpS 3.0 921 2763.0

HittingObstacle 3.0 661 1983.0

Jumping 3.0 746 2238.0

LyingDownFS 3.0 296 888.0

Running 3.0 1985 5955.0

SittingDown 3.0 200 600.0

StandingUpFL 3.0 216 648.0

StandingUpFS 3.0 153 459.0

Syncope 3.0 513 1539.0

Walking 3.0 1738 5214.0

smartphone accelerometer sampled at 50Hz was used

for recording, and the gravitational constant was re-

moved post-recording. The smartphone was placed in

equal parts in the left and right subjects’ pocket dur-

ing recording. The data is automatically segmented

into three-second windows around a magnitude peak.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

216

3 APPROACH

The effects of feature space reduction applied to

a HAR task are evaluated in the next sections us-

ing a person dependent shufﬂed and stratiﬁed 5-fold

cross-validation and a person independent leave-one-

person-out cross-validation. Each fold is evaluated

using a balanced accuracy (shown in Equation 1) as

the datasets are slightly imbalanced. The different

parameters are then compared using the average and

standard deviation over the folds. The baseline will

be optimized using a parameterwise grid search along

the recognizer stages using the best parameters ac-

cording to the independent evaluation for both the

next independent and dependent evaluations. Follow-

ing best practices, it starts with windowing and ends

with HMM parameters. The optimized baseline is

then compared against the best performing recognizer

with feature space reduction.

Let E donate the set of all activities a, then the

balanced or macro average accuracy is deﬁned as de-

scribed in Equation 1.

Balanced accuracy =

|E|

∑

Accuracy

(1)

The recognizer is set up similar to our previous

work, and the different stages are displayed in Figure

1. Each data segment is windowed, and features are

calculated on each window. The normalization then

removes the mean and scales the whole segment to

a standard deviation of one. An HMM with GMMs

modeling the emission probabilities performs the ﬁnal

classiﬁcation. The HMM is trained with the Expec-

tation Maximisation algorithm, and the GMMs use a

merge and split algorithm. For the reduction, the three

blue steps HMM, Stacking, and LDA are enabled. For

the UniMiB, a simple mirror is implemented (dis-

played in green) to offset the smartphones’ different

rotation between pockets. If the mean of the signal is

below zero, the signal is ﬂipped.

The blue HMM step denotes the force alignment

from the segmented sequential data to the topologies’

states. The alignment assigns each vector in the se-

quence the most likely HMM emission state. Force

alignments create a supervised vector-based problem

where each vector is assigned the activity and state

pair as a new target. Stacking refers to the process

of prepending the n previous feature vectors and ap-

pending the following n vectors to all vectors in a se-

quence, thereby increasing the time context and the

vector dimension by 2n + 1 the original dimension.

The LDA is then trained on the aligned and stacked

vectors and transforms the feature space along the

most discriminatory linear plane. During training, the

Data

Feature calculation

DecodingTraining

Normalization

HMM Stacking LDA

HMM

Axis Mirror

(UniMiB)

Windowing

Figure 1: Recognition pipeline. Green: unique steps for

UniMiB; Blue: feature space reduction steps.

HMM aligns the given segments and assigns each fea-

ture vector in the sequence its aligned activity and

state as a new target. During recognition, the activ-

ity of each segment is unknown. Therefore, the HMM

step is skipped, and the feature vectors are stacked

and transformed directly.

The recognizer was developed with SciPy (Virta-

nen et al., 2020), NumPy (Harris et al., 2020), scikit-

learn (Pedregosa et al., 2011), Matplotlib (Hunter,

2007), TSFEL (Barandas et al., 2020), and our in-

house decoder BioKIT (Telaar et al., 2014).

4 EVALUATION CSL19

First, a baseline for the CSL19 dataset is evaluated.

The HMM topology uses ﬁve states for each gait cy-

cle in an activity. Running, for example, contains

three full cycles and therefore uses ﬁfteen states. Sit,

stand, and the transitions between them are modeled

using a single state. For the initial hyperparameters,

a Hamming window with 20ms overlap, enabled nor-

malization, and ten HMM train-iterations are chosen

based on previous work (Hartmann et al., 2020). The

window size is determined in the ﬁrst experiment and

does not require an initial value. After optimiza-

tion, the best parameters are 100ms Hamming win-

dows with 50ms overlap, 10 HMM train-iterations,

and normalization enabled. The baseline achieves

a 93.7 ± 1.4% balanced accuracy in a person in-

dependent leave-one-person-out cross-validation and

97.8 ± 0.2% balanced accuracy in a stratiﬁed person

dependent 5-fold cross-validation.

The next step is to evaluate the best reduction pa-

rameters. The results of the experiment are displayed

in Figure 2. Note that the evaluation without reduc-

Feature Space Reduction for Human Activity Recognition based on Multi-channel Biosignals

217

Figure 2: Evaluation results for stacking and reduction on the CSL19 dataset. Left column: person independent evaluation,

right: person dependent evaluation. Upper row: the combination of stacking and reduction (dimension target, stacking

context), lower row: stacking alone. Without stacking, there is no reduction to 50 dimensions but a transformation of the

original 34 dimensions.

tion and a stacking context of zero corresponds to the

baseline recognizer. Figure 2 shows that stacking it-

self does not improve performance, except in the de-

pendent evaluation with a context of three. Reducing

to any dimension offsets this, but not above the base-

line. Notably, the recognition accuracy increases with

higher target dimensions independent of the stacking

context. Reducing to 50 dimensions performs better

than 30 and 20, independent of the speciﬁed stacking

context.

The best performance in an independent evalu-

ation is achieved using a stacking context of three

and reducing to a 50-dimensional feature space at

93.3 ± 3.9% accuracy. Notably, this is not higher

than the 93.7 ± 1.4% achieved without reduction. A

similar observation can be made in the person depen-

dent evaluation: 97.8±0.2% without and 97.4±0.3%

with the reduction. Figure 3 shows, that while the re-

duction based recognizer can better distinguish walk

and 90

◦

curves, standing is confused for sitting more

often, resulting in the same overall performance.

Additionally, it should be noted that the LDA

is trained with 238-dimensional feature vectors (17

channels times two features times seven vectors due

to context) on 370k samples. It distinguishes between

94 classes (ﬁve phases for most of the 22 activities).

The LDA of the best performing recognizer achieves a

27% accuracy between these classes in a 10-fold per-

son dependent evaluation.

5 EVALUATION UniMiB-SHAR

The activities are segmented with a ﬁxed-length peak

centered window rather than at the activity’s start and

Figure 3: Relative recognition accuracy improvements in

percentage points between baseline and reduction based

recognizer on the CSL19 dataset.

end. Therefore, the HMM topology is modeled with

a random state at the beginning and end of each ac-

tivity. The activities containing gait cycles are mod-

eled similarly to the CSL19 activities using ﬁve states

for each cycle. The falls are modeled with ten states,

as this performed best in a person independent cross-

validation.

The UniMiB-SHAR dataset contains fewer sam-

ples per segmented activity due to the lower sam-

pling rate. Therefore, a single grid search is feasi-

ble and executed. The best parameters are 400ms

Hamming windows with 320ms window overlap,

30 HMM train-iterations, and normalization enabled.

The baseline achieves a 59.7 ± 8.6% balanced ac-

curacy in a person independent leave-one-person-out

cross-validation.

The feature space reduction is evaluated on a grid

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

218

Figure 4: Evaluation results for stacking and reduction on the UniMiB dataset. Left column: person independent evaluation,

right: person dependent evaluation. Upper row: the combination of stacking and reduction (dimension target, stacking

context), lower row: stacking alone. Without stacking, there is no reduction to 20, 40, or 50 dimensions but a transformation

of the original six dimensions.

ranging from zero to six frames stacking context and

a target dimension from twenty to ﬁfty. The results

are displayed in Figure 4. Notably, stacking alone

already increases performance, and the reduction ex-

tends this. Similar to the CSL19 dataset, the target di-

mension inﬂuences performance more than the stack-

ing context. The best performing parameter combi-

nation with a target dimension of 30 and a context of

three performs at 69.5 ± 8.1% balanced accuracy in a

leave-one-person-out cross-validation. Compared to

the 59.7±8.6% accuracy in the baseline, this is an im-

provement by ten percentage points. This difference

is also apparent in the performance improvements for

each activity, as depicted in Figure 5.

Figure 5: Relative recognition accuracy improvements in

percentage points between baseline and reduction based

recognizer on the UniMiB dataset.

For easier comparison with previous work, the ac-

curacy for the best performing recognizer with fea-

ture space reduction was calculated in addition to the

balanced accuracy. The accuracy in the independent

evaluation is 77.0%, and the difference to the bal-

anced accuracy mainly arises from the ADLs occur-

ring more often and being recognized much better

than the falls.

Similar to the CSL19 dataset, the LDA performs

poorly when classifying the different vectors. The

LDA is trained with 42-dimensional feature vectors

(three channels, two features, seven vectors due to

context) on 360k samples. It distinguishes between

148 classes (ten phases for most of the 17 activities)

and achieves a 29% accuracy between these classes in

a 10-fold person dependent evaluation.

6 EVALUATION CSL18

The CSL18 dataset has been evaluated with and with-

out reduction previously, and a performance increase

with reduction could be shown (Hartmann et al.,

2020). The experiment is repeated with a balanced

accuracy metric to keep the results comparable with

the other two datasets. Furthermore, the CSL19 rec-

ognizer’s topology is applied to the CSL18 dataset in-

stead of the ﬁxed six-state topology used previously,

and the parameters are optimized with a person inde-

pendent evaluation rather than a dependent one. The

optimized baseline achieves 63.8±8.5% balanced ac-

curacy in an independent evaluation using the CSL19

topology, 30ms windows with 6ms overlap, and nor-

malization enabled.

Figure 6 shows the results from the stacking and

reduction experiments. In the person independent

evaluation, the best performances are achieved with-

Feature Space Reduction for Human Activity Recognition based on Multi-channel Biosignals

219

Figure 6: Evaluation results for stacking and reduction on the CSL18 dataset. Left column: person independent evaluation,

right: person dependent evaluation. Upper row: the combination of stacking and reduction (dimension target, stacking

context), lower row: stacking alone. Without stacking, there is no reduction to 50 dimensions but a transformation of the

original 40 dimensions.

out stacking at 70.6 ± 6% balanced accuracy when re-

ducing to a 20-dimensional feature space. A closer

look at the performance increases per activity shown

in Figure 7 reveals similar results to the CSL19

dataset with the walking derivatives being recognized

better, while the static activities stand and sit are con-

fused more often.

Figure 7: Relative recognition accuracy improvements in

percentage points between baseline and reduction based

recognizer on the CSL18 dataset.

7 DISCUSSION

There are several differences between the three rec-

ognizer and datasets, like the number of participants

(CSL18: 4, CSL19: 20, UniMiB: 30), the types of ac-

tivities (ADL, sport, falls), the HMM topology (gait-

phase based, ﬁxed number of states), or the evalu-

ated window length (30, 100, and 400ms). Across

all three, the feature space reduction with an LDA

did improve recognition accuracy for several activi-

ties. The most notable exception being the two only

static activities sit and stand. The resulting overall

performance is either similar or signiﬁcantly higher

compared to the baseline. Table 3 shows a summary

of the evaluation results.

The best performance on the CSL19 dataset with

93.7 ± 1.4% independent balanced accuracy ﬁts in

with the high ninety percent accuracies reported in

other works even though distinguishing more classes

(Rebelo et al., 2013) (Demrozi et al., 2020) (Lara

et al., 2012). The 92.8% person dependent accuracy

on the CSL18 dataset is slightly lower than the 94.9%

previously reported, as the parameters were optimized

with an independent evaluation instead of the purely

person dependent optimization done previously. The

results on the UniMiB dataset are directly comparable

to current research and better or on par. The person

independent balanced accuracy of 69.5± 8.1% is sig-

niﬁcantly higher than the 56.53% balanced accuracy

(Micucci et al., 2017), and the 77.0% independent ac-

curacy is on par with the 77.03% accuracy (Li et al.,

2018) previously reported.

The LDA does not perform as well as the HMM.

One reason might be the higher number of classes and

higher feature dimensionality compared to the HMM.

However, being trained with more than 350k samples

on both the CSL19 and UniMiB datasets, this seems

unlikely. Instead, this difference is probably caused

by the HMM modeling sequences and the LDA single

vectors, as discussed in section 1. Additionally, the

LDA is given samples that are very similar but differ-

ently labeled. Spin left with the left foot ﬁrst is shifted

half a gait cycle from spin left starting with the right

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

220

Table 3: Baseline and reduction results for both independent and dependent evaluation on each dataset.

Person independent Person dependent

Dataset Baseline Reduction Baseline Reduction

Balanced acc. Balanced acc. Accuracy Balanced acc. Balanced acc. Accuracy

CSL18 63.8 ± 8.5% 70.6 ± 6.0% 73.2% 86.7 ± 2.6% 92.1 ± 2.4% 92.8%

CSL19 93.7 ± 1.4% 93.3 ± 3.9% 93.6% 97.8 ± 0.2% 97.4 ± 0.3% 97.4%

UniMiB 59.7 ± 8.6% 69.5 ± 8.1% 77.0% 79.6 ± 0.5% 90.1 ± 0.7% 93.6%

foot. Nevertheless, the samples are assigned different

targets.

Despite these poor LDA discrimination perfor-

mances, its transformation does contribute to better

overall recognition performance, most notably in the

UniMiB and CSL18 datasets. The reduction was ben-

eﬁcial to the recognition of most activities, with the

notable exception of the two only static activities sit

and stand in the CSL18 and CSL19 dataset, which

are likely impacted by the normalization. The LDA’s

contribution should be investigated further in future

work, especially which activities can or cannot be im-

proved. Furthermore, the LDA should be compared

to a non-discriminatory reduction method like a PCA.

Non-discriminatory methods do not need to separate

targets and potentially handle overlapping classes bet-

ter.

These experiments show that the feature space

reduction using an LDA trained with HMM force

aligned targets can signiﬁcantly improve the recogni-

tion accuracy of activities as well as overall accuracy.

8 CONCLUSION AND FUTURE

WORK

The curse of dimensionality in HAR can be addressed

in several ways. For instance, through feature se-

lection or feature space transformations into lower-

dimensional spaces. The latter was evaluated on

the three HAR datasets CSL18, CSL19, and UniMiB

using force aligned labels and an LDA combined

with stacking. Initially, baselines were developed

for each dataset. In a person independent leave-one-

person-out cross-validation, the baselines achieved

93.7 ± 1.4% on the CSL19 and 59.7 ± 8.6% on the

UniMiB, and 63.8 ± 8.5% balanced accuracy on the

CSL18 dataset. Then the best context for stack-

ing and best feature dimension target were evaluated.

The reduction did not improve performance on the

CSL19 dataset. However, reduction increased perfor-

mance by ten percentage points to 69.5± 8.1% on the

UniMiB dataset and by seven points to 70.6 ± 6% on

the CSL18 dataset.

In future work, a closer investigation of the classes

as aligned by the HMM will be made and the beneﬁts

and downsides of the LDA-based reduction for differ-

ent activities closer investigated as well as compared

to a PCA-based reduction. Furthermore, experiments

with different feature selection methods will be con-

ducted, and their respective performance compared to

the LDA based reduction.

REFERENCES

Almaslukh, B., Artoli, A. M., and Al-Muhtadi, J.

(2018). A Robust Deep Learning Approach for

Position-Independent Smartphone-Based Human Ac-

tivity Recognition. Sensors, 18(11):3726.

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

Abreu, M., Bota, P., Liu, H., Schultz, T., and Gamboa,

H. (2020). TSFEL: Time Series Feature Extraction

Library. SoftwareX, 11:100456.

Bulling, A., Blanke, U., and Schiele, B. (2014). A tutorial

on human activity recognition using body-worn iner-

tial sensors. ACM Computing Surveys, 46(3):1–33.

Capela, N. A., Lemaire, E. D., and Baddour, N. (2015).

Feature selection for wearable smartphone-based hu-

man activity recognition with able bodied, elderly, and

stroke patients. PLoS ONE, 10(4):1–18.

Demrozi, F., Pravadelli, G., Bihorac, A., and Rashidi, P.

(2020). Human Activity Recognition using Inertial,

Physiological and Environmental Sensors: a Compre-

hensive Survey. 1(1).

Haeb-Umbach, R. and Ney, H. (1992). Linear discrimi-

nant analysis for improved large vocabulary contin-

uous speech recognition. In [Proceedings] ICASSP-

92: 1992 IEEE International Conference on Acous-

tics, Speech, and Signal Processing, volume 11, pages

13–16 vol.1. IEEE.

Harris, C. R., Millman, K. J., van der Walt, S. J., Gommers,

R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor,

J., Berg, S., Smith, N. J., Kern, R., Picus, M., Hoyer,

S., van Kerkwijk, M. H., Brett, M., Haldane, A., del

ıo, J. F., Wiebe, M., Peterson, P., G

erard-Marchant,

P., Sheppard, K., Reddy, T., Weckesser, W., Abbasi,

H., Gohlke, C., and Oliphant, T. E. (2020). Array pro-

gramming with NumPy. Nature, 585(7825):357–362.

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

reduction for multimodal human activity recognition.

In BIOSIGNALS 2020 - 13th International Confer-

ence on Bio-Inspired Systems and Signal Processing,

Proceedings; Part of 13th International Joint Confer-

Feature Space Reduction for Human Activity Recognition based on Multi-channel Biosignals

221

ence on Biomedical Engineering Systems and Tech-

nologies, BIOSTEC 2020.

Hu, H. and Zahorian, S. A. (2010). Dimensionality Reduc-

tion Methods for Hmm Phonetic Recognition. 2010

IEEE International Conference on Acoustics, Speech

and Signal Processing, pages 4854–4857.

Hunter, J. D. (2007). Matplotlib: A 2D Graphics Environ-

ment. Computing in Science & Engineering, 9(3):90–

95.

Lara,

O. D., Labrador, M. A., Lara, O. D., and Labrador,

M. A. (2012). A Survey on Human Activity Recogni-

tion using Wearable Sensors. IEEE Communications

Surveys & Tutorials, 15(3):1192–1209.

Li, F., Shirahama, K., Nisar, M. A., K

oping, L., and Grze-

gorzek, M. (2018). Comparison of feature learning

methods for human activity recognition using wear-

able sensors. Sensors (Switzerland), 18(2):1–22.

Liu, H. and Schultz, T. (2018). ASK: A Framework for Data

Acquisition and Activity Recognition. BIOSIGNALS

2018 - 11th International Conference on Bio-Inspired

Systems and Signal Processing, Proceedings; Part

of 11th International Joint Conference on Biomedi-

cal Engineering Systems and Technologies, BIOSTEC

2018, 4:262–268.

Liu, H. and Schultz, T. (2019). A Wearable Real-time Hu-

man Activity Recognition System using Biosensors

Integrated into a Knee Bandage. BIODEVICES 2019

- 12th International Conference on Biomedical Elec-

tronics and Devices, Proceedings; Part of 12th Inter-

national Joint Conference on Biomedical Engineering

Systems and Technologies, BIOSTEC 2019, pages 47–

55.

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

Recognition Using Acceleration Data from Smart-

phones.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,

Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,

Cournapeau, D., Brucher, M., Perrot, M., and Duch-

esnay,

E. (2011). Scikit-learn: Machine Learning

in Python. Journal of Machine Learning Research,

12(85):2825–2830.

Rebelo, D., Amma, C., Gamboa, H., and Schultz, T. (2013).

Human Activity Recognition for an Intelligent Knee

Orthosis. BIOSIGNALS 2013 - Proceedings of the In-

ternational Conference on Bio-Inspired Systems and

Signal Processing, pages 368–371.

Siohan, O. (1995). On the robustness of linear discrimi-

nant analysis as a preprocessing step for noisy speech

recognition. ICASSP, IEEE International Conference

on Acoustics, Speech and Signal Processing - Pro-

ceedings, 1(3):125–128.

Suto, J., Oniga, S., and Sitar, P. P. (2016). Comparison of

wrapper and ﬁlter feature selection algorithms on hu-

man activity recognition. In 2016 6th International

Conference on Computers Communications and Con-

trol (ICCCC), number Icccc, pages 124–129. IEEE.

Telaar, D., Wand, M., Gehrig, D., Putze, F., Amma, C.,

Heger, D., Vu, N. T., Erhardt, M., Schlippe, T., Janke,

M., Herff, C., and Schultz, T. (2014). BioKIT - Real-

time Decoder For Biosignal Processing. In The 15th

Annual Conference of the International Speech Com-

munication Association, Singapore, number Novem-

ber, pages 2650–2654.

Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M.,

Reddy, T., Cournapeau, D., Burovski, E., Peterson, P.,

Weckesser, W., Bright, J., van der Walt, S. J., Brett,

M., Wilson, J., Millman, K. J., Mayorov, N., Nel-

son, A. R., Jones, E., Kern, R., Larson, E., Carey,

C. J., Polat,

I., Feng, Y., Moore, E. W., VanderPlas,

J., Laxalde, D., Perktold, J., Cimrman, R., Henriksen,

I., Quintero, E. A., Harris, C. R., Archibald, A. M.,

Ribeiro, A. H., Pedregosa, F., van Mulbregt, P., Vi-

jaykumar, A., Bardelli, A. P., Rothberg, A., Hilboll,

A., Kloeckner, A., Scopatz, A., Lee, A., Rokem, A.,

Woods, C. N., Fulton, C., Masson, C., H

aggstr

om,

C., Fitzgerald, C., Nicholson, D. A., Hagen, D. R.,

Pasechnik, D. V., Olivetti, E., Martin, E., Wieser, E.,

Silva, F., Lenders, F., Wilhelm, F., Young, G., Price,

G. A., Ingold, G. L., Allen, G. E., Lee, G. R., Au-

dren, H., Probst, I., Dietrich, J. P., Silterra, J., Webber,

J. T., Slavi

c, J., Nothman, J., Buchner, J., Kulick, J.,

Sch

onberger, J. L., de Miranda Cardoso, J. V., Reimer,

J., Harrington, J., Rodr

ıguez, J. L. C., Nunez-Iglesias,

J., Kuczynski, J., Tritz, K., Thoma, M., Newville, M.,

ummerer, M., Bolingbroke, M., Tartre, M., Pak, M.,

Smith, N. J., Nowaczyk, N., Shebanov, N., Pavlyk,

O., Brodtkorb, P. A., Lee, P., McGibbon, R. T., Feld-

bauer, R., Lewis, S., Tygier, S., Sievert, S., Vigna, S.,

Peterson, S., More, S., Pudlik, T., Oshima, T., Pin-

gel, T. J., Robitaille, T. P., Spura, T., Jones, T. R.,

Cera, T., Leslie, T., Zito, T., Krauss, T., Upadhyay,

U., Halchenko, Y. O., and V

azquez-Baeza, Y. (2020).

SciPy 1.0: fundamental algorithms for scientiﬁc com-

puting in Python. Nature Methods, 17(3):261–272.

Weiner, J. and Schultz, T. (2018). Selecting Features for

Automatic Screening for Dementia Based on Speech.

pages 747–756.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

222