Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 41 of the 42 references it cites were found in a registry or an index, and 1 are not checked yet; 6 of the 43 sources its claims cite were checked for support.

Comparison of Feature Learning Methods for Human Activity Recognition Using Wearable Sensors

Open access

Abstract

Getting a good feature representation of data is paramount for Human Activity Recognition (HAR) using wearable sensors. An increasing number of feature learning approaches—in particular deep-learning based—have been proposed to extract an effective feature representation by analyzing large amounts of data. However, getting an objective interpretation of their performances faces two problems: the lack of a baseline evaluation setup, which makes a strict comparison between them impossible, and the insufficiency of implementation details, which can hinder their use. In this paper, we attempt to address both issues: we firstly propose an evaluation framework allowing a rigorous comparison of features extracted by different methods, and use it to carry out extensive experiments with state-of-the-art feature learning approaches. We then provide all the codes and implementation details to make both the reproduction of the results reported in this paper and the re-use of our framework easier for other researchers. Our studies carried out on the OPPORTUNITY and UniMiB-SHAR datasets highlight the effectiveness of hybrid deep-learning architectures involving convolutional and Long-Short-Term-Memory (LSTM) to obtain features characterising both short- and long-term time dependencies in the data.

Verified 2026-10-05 · 1 reference not checked yet

Overview

CoChat Summary

TL;DR

The paper proposes an evaluation framework that fixes all stages of the Activity Recognition Chain except feature extraction, then compares seven feature learning methods on OPPORTUNITY and UniMiB-SHAR. Hybrid CNN-LSTM architectures are the most robust across datasets, while codebook soft assignment is especially effective on UniMiB-SHAR and feature fusion improves results.

Key findings

  • On OPPORTUNITY, the hybrid CNN-LSTM model was best, with a reported weighted F1-score of 70.86%; LSTM’s average F1-score exceeded CNN’s by 4.45 percentage points.
  • On UniMiB-SHAR, codebook soft assignment (CBS) achieved the highest accuracy and weighted F1-score at 77.03% and 75.93%, respectively, while CNN obtained the best average F1-score at 64.65%.
  • Hand-crafted features underperformed the best methods by 7.39% average F1-score on OPPORTUNITY and 49.93% on UniMiB-SHAR.
  • Fusing codebook and hybrid CNN-LSTM features improved performance on both datasets; changing the number of sensor channels or sliding-window size did not change the overall ranking of methods, with T=64 best in the reference OPPORTUNITY setup.

Methods

  • The evaluation framework fixes all ARC stages except feature extraction and compares seven approaches: hand-crafted features, MLP, CNN, LSTM, hybrid CNN-LSTM, autoencoder, and codebook (hard and soft assignment), using a linear soft-margin SVM classifier and metrics including accuracy, weighted F1-score, and average F1-score.
  • OPPORTUNITY: 4 subjects, 18 classes including NULL, 30 Hz, 145 sensor channels reduced to 107, sliding window T=64 (2 s) with stride 3; training used ADL1–ADL3 and Drill (211,265 frames), testing used ADL4–ADL5 (78,029 frames). UniMiB-SHAR: 30 subjects, 17 classes, 3D accelerometer at 50 Hz, T=151 (about 3 s), 11,771 windows, evaluated with 30-fold leave-one-subject-out cross-validation.
  • Deep models were implemented in Keras/TensorFlow, trained with ADADELTA for 50 epochs, and fine-tuned by removing the softmax layer and using pre-trained models to compute features for SVM training; hyper-parameters were selected by manual search.

Limitations

  • Stated: strict comparison with previous works is not possible because of differing evaluation frameworks; hyper-parameters were manually searched rather than optimized; on UniMiB-SHAR, the codebook was constructed using subsequences from all 30 subjects, including test subjects, which might slightly skew CB performances; larger time windows may contain multiple activities, making majority labeling less accurate; UniMiB-SHAR’s energy-based segmentation lacks temporal consistency and transition frames, possibly explaining LSTM’s relatively poor performance.
  • Evident from the text: the comparison covers only two benchmark datasets, and the reported rankings depend on the specific segmentation type and dataset characteristics.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/s18020679

Cites

41 of 42 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Activity Recognition Chain (ARC). Raw data are firstly acquired from sensors. After pre-processing, segments of data are extracted (Segmentation) and values relevant to the recognition problem are computed from them (Feature extraction). A classifier is then trained and evaluated using those features (Classification). In our framework, all steps of the ARC—except the feature extraction part—are fixed, and a Support-Vector-Machine (SVM) classifier is used for the classification stage.
Figure 1 Activity Recognition Chain (ARC). Raw data are firstly acquired from sensors. After pre-processing, segments of data are extracted (Segmentation) and values relevant to the recognition problem are computed from them (Feature extraction). A classifier is then trained and evaluated using those features (Classification). In our framework, all steps of the ARC—except the feature extraction part—are fixed, and a Support-Vector-Machine (SVM) classifier is used for the classification stage.CC BY · PubMed Central
Figure 2 An overview of the codebook approach.
Figure 2 An overview of the codebook approach.CC BY · PubMed Central
Figure 3 Architecture of a MLP model with H hidden layers for sensor-based HAR. Input data from the different sensor channels are first flattened into a (T×S)-dimensional vector and then fed to the hidden layers. All layers are fully-connected. The numbers in parenthesis indicate the number of neurons per layer. T, S and N designate the time length of the input data, the number of sensor channels, and the number of classes, respectively.
Figure 3 Architecture of a MLP model with H hidden layers for sensor-based HAR. Input data from the different sensor channels are first flattened into a (T×S)-dimensional vector and then fed to the hidden layers. All layers are fully-connected. The numbers in parenthesis indicate the number of neurons per layer. T, S and N designate the time length of the input data, the number of sensor channels, and the number of classes, respectively.CC BY · PubMed Central
Figure 4 Architecture of a CNN model for sensor-based HAR. T, S and nk designate the time length of the input data, the number of sensor channels, and the number of convolutional kernels of the kth layer, respectively. The convolutional and pooling kernels of the kth layer process patches of the input data of sizes (fT^(k),fS^(k)) and (pT^(k),pS^(k)), respectively. Neurons of intermediate convolutional layers perform convolution products across all convolutional maps of the previous layer.
Figure 4 Architecture of a CNN model for sensor-based HAR. T, S and nk designate the time length of the input data, the number of sensor channels, and the number of convolutional kernels of the kth layer, respectively. The convolutional and pooling kernels of the kth layer process patches of the input data of sizes (fT^(k),fS^(k)) and (pT^(k),pS^(k)), respectively. Neurons of intermediate convolutional layers perform convolution products across all convolutional maps of the previous layer.CC BY · PubMed Central
Figure 5 Architecture of a LSTM network for sensor-based HAR (left). The intermediate and last LSTM layers are organized in a many-to-many and many-to-one layout, respectively. Each LSTM layer is composed of LSTM cells (right). xt, mt and ct represent the cell input, memory, and output at time t, respectively. In addition, ⨁ and ⨂ refer to the element-wise addition and multiplication, respectively. W*, U*, b* and σ* designate internal weight matrices, bias vector and activation function (for gate *∈{i,o,f}). σ1 and σ2 are internal activation functions applied on the input and memory of the cell, respectively.
Figure 5 Architecture of a LSTM network for sensor-based HAR (left). The intermediate and last LSTM layers are organized in a many-to-many and many-to-one layout, respectively. Each LSTM layer is composed of LSTM cells (right). xt, mt and ct represent the cell input, memory, and output at time t, respectively. In addition, ⨁ and ⨂ refer to the element-wise addition and multiplication, respectively. W*, U*, b* and σ* designate internal weight matrices, bias vector and activation function (for gate *∈{i,o,f}). σ1 and σ2 are internal activation functions applied on the input and memory of the cell, respectively.CC BY · PubMed Central
Figure 6 Architecture of a hybrid CNN+LSTM model for sensor-based HAR. Each slice along the time dimension of the output of the convolutional block(s) (in blue) is fed to one LSTM cell. All LSTM layers are organized in a many-to-many pattern, except the last which follows a many-to-one scheme.
Figure 6 Architecture of a hybrid CNN+LSTM model for sensor-based HAR. Each slice along the time dimension of the output of the convolutional block(s) (in blue) is fed to one LSTM cell. All LSTM layers are organized in a many-to-many pattern, except the last which follows a many-to-one scheme.CC BY · PubMed Central
Figure 7 Architecture of an autoencoder model for sensor-based HAR. The numbers in parenthesis refer to the number of neurons per layer. H denotes the number of hidden layers in both the encoder and decoder.
Figure 7 Architecture of an autoencoder model for sensor-based HAR. The numbers in parenthesis refer to the number of neurons per layer. H denotes the number of hidden layers in both the encoder and decoder.CC BY · PubMed Central
Figure 8 Average F1-scores of the feature learning methods on the OPPORTUNITY dataset, using different numbers of sensor channels (ranked by decreasing variance).
Figure 8 Average F1-scores of the feature learning methods on the OPPORTUNITY dataset, using different numbers of sensor channels (ranked by decreasing variance).CC BY · PubMed Central

Verification

Claims the paper makes

6 Valid · 2 Not valid · 25 Not checked · 75 Own results and interpretations

53 more claims

Create a free account to see all 108 claims and to verify the 37 references still open.

See all claims — free account
  • Probably not a substantial claim. 35
  • It may state two separate claims as one. 23
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

1 of 1

Citations

Cites

Show all 41 references

Research works printed but not linked: 2

  1. Cook, D.J., and Krishnan, N.C. (2015). Activity Learning: Discovering, Recognizing, and Predicting Human Behaviour from Sensor Data, John Wiley & Sons.
  2. Han, J., Kamber, M., and Pei, J. (2011). Data Mining: Concepts and Techniques, Morgan Kaufmann.

Cited by

See all 323

Metadata

Authors
Li, F., Shirahama, K., Nisar, M., Köping, L., & Grzegorzek, M.
Affiliations
  • University of Siegen (DE)
  • University of Economics in Katowice (PL)
Corresponding author
Li, Frédéric
Venue
Sensors
Publisher
MDPI AG
Published
2018-02-24
Volume
18
Issue
2
Pages
679
ISSN
1424-8220 (ISSN-L)
OpenAlex
W2789868604
Cited by
302
References
41
Citations by year
2026: 4 · 2025: 27 · 2024: 36 · 2023: 42 · 2022: 43 · 2021: 54 · 2020: 49 · 2019: 34 · 2018: 13
FWCI
18.76
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Bundesministerium für Bildung und Forschung (16SV7223K, 16SV7512)
  • Deutsche Forschungsgemeinschaft (GRK1564)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Engineering › Biomedical Engineering › Gait Recognition and Analysis
  • Physical Sciences › Computer Science › Signal Processing › Time Series Analysis and Forecasting
Keywords
human activity recognition, wearable sensors, representation learning, feature representation, evaluation framework, deep learning, convolutional neural networks, long short-term memory, CNN-LSTM hybrid model, Opportunity dataset, time dependence
MeSH terms
Machine Learning, Wearable Electronic Devices, Humans, Human Activities, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Comparison of Feature Learning Methods for Human Activity Recognition Using Wearable Sensors. https://app.cochat.ai/papers/10.3390/s18020679
The paper
Li, F., Shirahama, K., Nisar, M., Köping, L., & Grzegorzek, M. (2018). Comparison of Feature Learning Methods for Human Activity Recognition Using Wearable Sensors. Sensors. https://doi.org/10.3390/s18020679

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.