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

The DOI matches its Crossref record; 33 of the 33 references it cites were found in a registry or an index; 5 of the 23 sources its claims cite were checked for support.

Deep Recurrent Neural Networks for Human Activity Recognition

Open access

Abstract

Adopting deep learning methods for human activity recognition has been effective in extracting discriminative features from raw input sequences acquired from body-worn sensors. Although human movements are encoded in a sequence of successive samples in time, typical machine learning methods perform recognition tasks without exploiting the temporal correlations between input data samples. Convolutional neural networks (CNNs) address this issue by using convolutions across a one-dimensional temporal sequence to capture dependencies among input data. However, the size of convolutional kernels restricts the captured range of dependencies between data samples. As a result, typical models are unadaptable to a wide range of activity-recognition configurations and require fixed-length input windows. In this paper, we propose the use of deep recurrent neural networks (DRNNs) for building recognition models that are capable of capturing long-range dependencies in variable-length input sequences. We present unidirectional, bidirectional, and cascaded architectures based on long short-term memory (LSTM) DRNNs and evaluate their effectiveness on miscellaneous benchmark datasets. Experimental results show that our proposed models outperform methods employing conventional machine learning, such as support vector machine (SVM) and k-nearest neighbors (KNN). Additionally, the proposed models yield better performance than other deep learning techniques, such as deep believe networks (DBNs) and CNNs.

Overview

CoChat Summary

TL;DR

The paper proposes LSTM-based deep recurrent neural network (DRNN) architectures—unidirectional, bidirectional, and cascaded—for human activity recognition from wearable sensor sequences, including variable-length windows. Across benchmark datasets, the proposed models reportedly outperform conventional machine learning and other deep learning baselines.

Key findings

  • On UCI-HAD, the unidirectional DRNN with four layers achieved 96.7% overall accuracy, outperforming CNNs, SVM, and sequential ELM.
  • On Skoda, the cascaded DRNN with one bidirectional layer and two upper unidirectional layers achieved 92.6% overall accuracy, outperforming HMMs, DBNs, and CNNs.
  • On Opportunity, the bidirectional DRNN with three layers outperformed DBNs, SVM, CNNs, and a CNN + unidirectional RNN state-of-the-art method, using F1 score because the dataset is imbalanced.
  • On Daphnet FOG, the cascaded DRNN outperformed KNN and CNNs in F1 score; on USC-HAD, the unidirectional four-layer model performed best versus CNNs, LS-SVM, and random forest, though no exact accuracy is given in the text.

Methods

  • Three LSTM-based DRNN architectures were developed: unidirectional, bidirectional, and cascaded (bidirectional first layer plus unidirectional upper layers), with late fusion using the sum rule over per-timestep scores.
  • Models were evaluated on five public HAR datasets: UCI-HAD, USC-HAD, Opportunity, Daphnet FOG, and Skoda; training used 80% of data and testing used 20%, raw data without further preprocessing, Adam optimization, mean cross entropy, dropout, and a GPU-based TensorFlow framework.
  • Performance metrics were precision, recall, accuracy, and F1-score.

Limitations

  • Stated: future work includes large-scale and complex human activities, transfer learning between diverse datasets, and resource-efficient DRNN implementation for low-power devices.
  • Stated: USC-HAD elevator up/down classes were combined because the model could not differentiate them using only a single IMU; barometer readings would be needed. Daphnet FOG had low recall and precision for the “Freeze” class due to dominance of the “Normal” class, and Opportunity is imbalanced due to the Null class.
  • Evident: several dataset-specific results are reported only through figures/confusion matrices without exact accuracy or F1 values in the text, and the paper inconsistently refers to four datasets in the Discussion/Conclusions while the experimental setup describes five.

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

Integrity signals

Registry

Crossref · 10.3390/s17112556

Cites

33 of 33 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Schematic diagram of an RNN node where ht−1 is the previous hidden state, xt is the current input sample, ht is the current hidden state, yt is the current output, and ℱ is the activation function.
Figure 1 Schematic diagram of an RNN node where ht−1 is the previous hidden state, xt is the current input sample, ht is the current hidden state, yt is the current output, and ℱ is the activation function.CC BY · PubMed Central
Figure 2 Schematic of LSTM cell structure with an internal recurrence ct and an outer recurrence ht. Cell gates are the input gate it, input modulation gate gt, forget gate ft, and output gate ot. In contrast to an RNN node, the current output 𝓎t is considered equal to current hidden state ht.
Figure 2 Schematic of LSTM cell structure with an internal recurrence ct and an outer recurrence ht. Cell gates are the input gate it, input modulation gate gt, forget gate ft, and output gate ot. In contrast to an RNN node, the current output 𝓎t is considered equal to current hidden state ht.CC BY · PubMed Central
Figure 3 The proposed HAR architecture. The inputs are raw signals obtained from multimodal-sensors, segmented into windows of length T and fed into LSTM-based DRNN model. The model outputs class prediction scores for each timestep, which are then merged via late-fusion and fed into the softmax layer to determine class membership probability.
Figure 3 The proposed HAR architecture. The inputs are raw signals obtained from multimodal-sensors, segmented into windows of length T and fed into LSTM-based DRNN model. The model outputs class prediction scores for each timestep, which are then merged via late-fusion and fed into the softmax layer to determine class membership probability.CC BY · PubMed Central
Figure 4 Unidirectional LSTM-based DRNN model consisting of an input layer, several hidden layers, and an output layer. The number of hidden layers is a hyperparameter that is tuned during training.
Figure 4 Unidirectional LSTM-based DRNN model consisting of an input layer, several hidden layers, and an output layer. The number of hidden layers is a hyperparameter that is tuned during training.CC BY · PubMed Central
Figure 5 Bidirectional LSTM-based DRNN model consisting of an input layer, multiple hidden layers, and an output layer. Every layer has a forward LSTM^fℓ and a backward LSTM^bℓ track, and the number of hidden layers is a hyperparameter that is tuned during training.
Figure 5 Bidirectional LSTM-based DRNN model consisting of an input layer, multiple hidden layers, and an output layer. Every layer has a forward LSTM^fℓ and a backward LSTM^bℓ track, and the number of hidden layers is a hyperparameter that is tuned during training.CC BY · PubMed Central
Figure 6 Cascaded unidirectional and bidirectional LSTM-based DRNN model. The first layer is bidirectional, whereas the upper layers are unidirectional. The number of hidden unidirectional layers is a hyperparameter that is tuned during training.
Figure 6 Cascaded unidirectional and bidirectional LSTM-based DRNN model. The first layer is bidirectional, whereas the upper layers are unidirectional. The number of hidden unidirectional layers is a hyperparameter that is tuned during training.CC BY · PubMed Central
Figure 7 Accuracy and cost of the unidirectional DRNN model for the USC-HAD dataset over mini-batch training iterations: (a) training and testing accuracies; (b) cross-entropy cost between ground truth labels and predicted labels for both training and testing.
Figure 7 Accuracy and cost of the unidirectional DRNN model for the USC-HAD dataset over mini-batch training iterations: (a) training and testing accuracies; (b) cross-entropy cost between ground truth labels and predicted labels for both training and testing.CC BY · PubMed Central
Figure 8 Performance results of the proposed unidirectional DRNN model for the UCI-HAD dataset: (a) Confusion matrix for the test set containing the activity recognition results. The rows represent the true labels and the columns represent the model classification results; (b) Comparative accuracy of the proposed model against other methods.
Figure 8 Performance results of the proposed unidirectional DRNN model for the UCI-HAD dataset: (a) Confusion matrix for the test set containing the activity recognition results. The rows represent the true labels and the columns represent the model classification results; (b) Comparative accuracy of the proposed model against other methods.CC BY · PubMed Central
Figure 9 Performance results of the proposed unidirectional DRNN model for USC-HAD dataset: (a) Confusion matrix for the test set displaying activity recognition results with per-class precision and recall; (b) Comparative accuracy of proposed model against other methods.
Figure 9 Performance results of the proposed unidirectional DRNN model for USC-HAD dataset: (a) Confusion matrix for the test set displaying activity recognition results with per-class precision and recall; (b) Comparative accuracy of proposed model against other methods.CC BY · PubMed Central
Figure 10 Performance results of the proposed bidirectional DRNN model for the Opportunity dataset: (a) Confusion matrix for the test set as well as per-class precision and recall results; (b) Comparative F1 score of proposed model against other methods.
Figure 10 Performance results of the proposed bidirectional DRNN model for the Opportunity dataset: (a) Confusion matrix for the test set as well as per-class precision and recall results; (b) Comparative F1 score of proposed model against other methods.CC BY · PubMed Central
Figure 11 Performance results of the proposed cascaded DRNN model for the Daphnet FOG dataset: (a) Confusion matrix for the test set, along with per-class precision and recall; (b) F1 score of the proposed method in comparison with other methods.
Figure 11 Performance results of the proposed cascaded DRNN model for the Daphnet FOG dataset: (a) Confusion matrix for the test set, along with per-class precision and recall; (b) F1 score of the proposed method in comparison with other methods.CC BY · PubMed Central
Figure 12 Performance results of the proposed cascaded DRNN model for the Skoda dataset: (a) Confusion matrix for the test set as well as per-class precision and recall results; (b) Comparative accuracy of proposed model against other methods.
Figure 12 Performance results of the proposed cascaded DRNN model for the Skoda dataset: (a) Confusion matrix for the test set as well as per-class precision and recall results; (b) Comparative accuracy of proposed model against other methods.CC BY · PubMed Central

Verification

Claims the paper makes

3 Valid · 0 Not valid · 10 Not checked · 27 Own results and interpretations

23 more claims

Create a free account to see all 40 claims and to verify the 18 references still open.

See all claims — free account
  • Probably not a substantial claim. 4
  • It may state two separate claims as one. 7
  • The quote only partly states this claim. 2

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.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Research works printed but not linked: 1

  1. Goodfellow, I., Bengio, Y., and Courville, A. (2016). Optimization for Training Deep Models. Deep Learning, The MIT Press.

Cited by

See all 545

Metadata

Authors
Murad, A., & Pyun, J.-Y.
Affiliations
Chosun University (KR)
Corresponding author
Pyun, Jae-Young
Venue
Sensors
Publisher
MDPI AG
Published
2017-11-06
Volume
17
Issue
11
Pages
2556
ISSN
1424-8220 (ISSN-L)
OpenAlex
W2767979715
Cited by
524
References
33
Citations by year
2026: 27 · 2025: 45 · 2024: 57 · 2023: 60 · 2022: 81 · 2021: 82 · 2020: 80 · 2019: 45 · 2018: 19 · 2017: 2
FWCI
16.57
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
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 › Computer Vision and Pattern Recognition › Human Pose and Action Recognition
Keywords
human activity recognition, deep recurrent neural networks, long short-term memory, wearable sensors, temporal dependencies, bidirectional LSTM, wearable sensor data
MeSH terms
Machine Learning, Algorithms, 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 (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

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.