CoChat
Sign Up
Cited by 333Context-Aware Activity Recognition SystemsGait Recognition and AnalysisHuman Pose and Action Recognition

InnoHAR: A Deep Neural Network for Complex Human Activity Recognition

Open access
Xu, Cheng, Chai, Duo, He, Jie, Zhang, Xiaotong, Duan, Shihong·2019·IEEE Access
DOI 10.1109/access.2018.2890675
Open this paper in CoChat — free account

Abstract

Human activity recognition (HAR) based on sensor networks is an important research direction in the fields of pervasive computing and body area network. Existing researches often use statistical machine learning methods to manually extract and construct features of different motions. However, in the face of extremely fast-growing waveform data with no obvious laws, the traditional feature engineering methods are becoming more and more incapable. With the development of deep learning technology, we do not need to manually extract features and can improve the performance in complex human activity recognition problems. By migrating deep neural network experience in image recognition, we propose a deep learning model (InnoHAR) based on the combination of inception neural network and recurrent neural network. The model inputs the waveform data of multi-channel sensors end-to-end. Multi-dimensional features are extracted by inception-like modules by using various kernel-based convolution layers. Combined with GRU, modeling for time series features is realized, making full use of data characteristics to complete classification tasks. Through experimental verification on three most widely used public HAR datasets, our proposed method shows consistent superior performance and has good generalization performance, when compared with the state-of-the-art.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes InnoHAR, a deep model combining Inception-like modules with GRU layers for wearable-sensor human activity recognition. It reports higher F-measure than CNN and DeepConvLSTM baselines on three public HAR datasets and demonstrates real-time prediction on an embedded board.

Key findings

  • InnoHAR achieved the best F-measure on all three datasets: 0.946 on Opportunity, 0.935 on PAMAP2, and 0.945 on Smartphones, compared with CNN (0.851, 0.88, 0.915) and DeepConvLSTM (0.915, 0.905, 0.91).
  • On Opportunity, the paper reports about a 9% increase over CNN and a 3% improvement over DeepConvLSTM; on PAMAP2, nearly a 5% gap between best and worst and a 3% advantage over DeepConvLSTM.
  • Adding an extra CNN layer to DeepConvLSTM did not improve accuracy and caused a decline; replacing LSTM with bi-LSTM increased accuracy slightly but was difficult to parallelize and time-consuming.
  • On a MinnowBoard Turbot Dual Core Board, the model took 152.02 s to predict all 9894 sliding windows, reaching 65.09 pieces/s, which the authors say easily supports real-time activity prediction for data recorded at 30 Hz.

Methods

  • InnoHAR uses four Inception-like modules with 1×1, 1×3, and 1×5 convolution kernels plus pooling, followed by two GRU layers; ELU is used as the nonlinear activation, and multi-channel waveform data are input end-to-end.
  • Experiments use three public datasets: Opportunity (113 on-body channels, 30 Hz, 18 classes including Null, 24-line sliding window sliding 12 lines, 9,984 test pieces), PAMAP2 (9 participants, 18 activities, 52 dimensions, accelerometer downsampled to 33.3 Hz, 5.12 s windows with 1 s stepping), and Smartphone (30 subjects, six activities, 50 Hz, 2.56 s windows with 50% overlap).
  • Performance is evaluated by F1-score with class weighting; the model is compared with CNN and DeepConvLSTM, modified state-of-the-art variants, and tested for efficiency on an embedded MinnowBoard system.

Limitations

  • The paper states future work will continue adjusting the network structure, including kernel size and connection method, and will further explore data imbalance in real-life human activity recognition.
  • Evident from the text: evaluation is limited to three public datasets, with no statistical significance testing or broad ablation beyond modified DeepConvLSTM variants.
  • Evident from the text: the reported test-window count differs between sections (9,984 pieces for testing vs. 9894 sliding windows in the efficiency analysis), and the PAMAP2 window description says “non-overlapping” while also stating 78% overlap.

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

Integrity signals

Registry

Crossref · 10.1109/access.2018.2890675

Cites

36 of 36 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 23 cited sources checked · 0 supported · 0 flagged

  • Findingnone given
    § B. COMPARED WITH IMPROVED STATE-OF-THE-ARTS · p. 8

    Transforming DeepConvLSTM into a bidirectional LSTM increases accuracy slightly.

    “In contrast, when DeepConvLSTM was transformed into a bidirectional LSTM, the accuracy did increase slightly.”
    Quote in the paperNot a substantial claim
  • Findingnone given
    § D. PERFORMANCE MEASURE · p. 6

    The gestures of the OPPORTUNITY dataset are extremely imbalanced, with the Null class accounting for more than 70% of all the data.

    “The gestures of OPPORTUNITY dataset are extremely imbalanced, the Null class accounts for more than 70% of all the data.”
    Quote in the paperNot a substantial claim
  • Interpretationreasoning
    § I. INTRODUCTION · p. 2

    This concatenation greatly reduces interference and misidentification caused by no obvious segmentation.

    “which greatly reduces the interference and misidenti cation caused by no obvious segmentation.”
    Quote in the paperSubstantial claim
  • Premisecitation
    § A. CONVOLUTIONAL NEURAL NETWORK · p. 3

    Extracting more useful features from the original signal requires sufficient expert knowledge, which inevitably limits a systematic exploration of the feature space.

    “Extracting more useful features from the original signal requires suf cient expert knowledge, which inevitably limits a systematic exploration of the feature space [17].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] source

      Not checked in the free run.

  • Premisecitation
    § B. TEMPORAL FEATURES EXTRACTION · p. 4

    Vanilla RNN has a problem of gradient disappearance.

    “Experience has shown that vanilla RNN has a problem of gradient disappearance [18]”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [18]

      Not checked in the free run.

  • May combine two claims. 24
  • Classified as probably not a substantial claim. 11
  • The quote only partly states this claim. 2

  • May combine two claims.

    Findingown data
    § CHENG XU 1,2 , (Student Member, IEEE), DUO CHAI 1,2 , JIE HE 1,2 , XIAOTONG ZHANG 1,2 , (Senior Member, IEEE), AND SHIHONG DUAN 1,2 · p. 1

    The proposed InnoHAR method shows consistent superior performance and good generalization on three widely used public HAR datasets compared with state-of-the-art methods.

    “Through experimental veri cation on three most widely used public HAR datasets, our proposed method shows consistent superior performance and has good generalization performance, when compared with the state-of-the-art.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The three public HAR datasets used are the most widely used public HAR datasets.

  • May combine two claims.

    Findingown data
    § I. INTRODUCTION · p. 2

    Blindly superposing network layers and accumulating parameters multiplies training and verification time and reduces the efficiency of learning and prediction.

    “By comparison experiments on three most widely used public datasets, it is proved that the blindly superposition of network layers and the accumulation of parameters, in addition to multiplying the training and veri cation time, reduces the ef ciency of learning and prediction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning
    § I. INTRODUCTION · p. 2

    Blindly superposing network layers and accumulating parameters cannot improve recognition accuracy and leads to a decline in prediction results.

    “It cannot improve the accuracy of recognition, but lead to a decline in prediction results, perhaps because the network is too deep and insuf cient data volume makes the structure dif cult to be fully trained.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 8

    On the Smartphone dataset, InnoHAR is superior to both CNN and DeepConvLSTM.

    “For smartphones dataset, InnoHAR is also superior to the other two methods and maintains good generalization performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 8

    On the Opportunity dataset, InnoHAR improves F1-score by about 9% over CNN and by 3% over DeepConvLSTM.

    “For Opportunit dataset, our proposed model has a signi cant increase of about 9% compared with the CNN model of Yang et al. [25], as well as a 3% improvement compared with DeepConvLSTM.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 8

    On the PAMAP2 dataset, InnoHAR improves F1-score by nearly 5% over CNN and by 3% over DeepConvLSTM.

    “For PAMAP2 dataset, we can also observe nearly a 5% gap between the best (proposed model) and worst (CNN [25]), and InnoHAR also has a 3% performance advantage over DeepConvLSTM.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] source

      Not checked in the free run.

  • May combine two claims.

    Findingcitation
    § V. RESULTS AND ANALYSIS · p. 7

    DeepConvLSTM achieves an F1-score of 0.915 on the Opportunity dataset, 0.905 on the PAMAP2 dataset, and 0.91 on the Smartphone dataset.

    “| DeepConvLSTM [19] | 0.915 | 0.905 | 0.91 |”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [19] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data
    § V. RESULTS AND ANALYSIS · p. 7

    InnoHAR achieves an F1-score of 0.946 on the Opportunity dataset, 0.935 on the PAMAP2 dataset, and 0.945 on the Smartphone dataset.

    “| Proposed method (InnoHAR) | 0.946 | 0.935 | 0.945 |”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 8

    InnoHAR outperforms both CNN and DeepConvLSTM on the Opportunity, PAMAP2, and Smartphone datasets.

    “In terms of overall performance, the experimental results show that our proposed InnoHAR model based on Inception-like module has better performance than both CNN [25] and DeepConvLSTM [19].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [25] source

      Not checked in the free run.

    • cites [19] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data
    § B. COMPARED WITH IMPROVED STATE-OF-THE-ARTS · p. 8

    Adding an additional CNN layer to DeepConvLSTM to increase depth does not improve recognition accuracy and causes a certain degree of decline.

    “In Fig. 7, we can see that when an additional CNN layer is added to simply increase the depth of model, the recognition accuracy model is not improved as expected, but has a certain degree of decline.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation
    § V. RESULTS AND ANALYSIS · p. 7

    CNN achieves an F1-score of 0.851 on the Opportunity dataset, 0.88 on the PAMAP2 dataset, and 0.915 on the Smartphone dataset.

    “| CNN [25] | 0.851 | 0.88 | 0.915 |”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given
    § B. TEMPORAL FEATURES EXTRACTION · p. 4

    The entire InnoHAR network structure delivers satisfactory results in both predictive performance and time efficiency.

    “The entire network structure delivers satisfactory results in both predictive performance and time ef ciency.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: GRU performs better than LSTM in terms of time efficiency.
    • Rests on an assumption: GRU and LSTM show better performance in dealing with long sequences based problems.

  • May combine two claims. The quote only partly states this claim.

    Findingown data
    § C. EFFICIENCY ANALYSIS · p. 8

    InnoHAR took 152.02 s to predict all 9894 sliding windows on the MinnowBoard, reaching a predicted speed of 65.09 pieces/s.

    “In the prediction of all previous test data, our model took 152.02 s and completed the prediction of all 9894 sliding windows. The predicted speed reached 65.09 pieces/s.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 8

    InnoHAR maintains consistent superior performance and good generalization on different public datasets, including the three most widely used public datasets.

    “To be more speci c, Fig. 6 shows the confusion matrix of proposed method and two state-of-the-arts in one test run, from which we can see that proposed method maintains consistent superior performance on different public datasets and has good generalization performance in the recognition of complex human activities.”
    • Also stated in: § VI. CONCLUSION · p. 9 · plain
      “Compared with baselines and state-of-the-arts, our proposed method shows consistent superior performance and has good generalization performance on three most widely used public datasets.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The Opportunity, PAMAP2, and Smartphone datasets are representative of real-world complex human activity recognition, so performance on them generalizes to real-world HAR.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given
    § C. GOOGLENET AND INCEPTION MODULE · p. 4

    Blindly increasing network size causes an exponential increase in computing resource requirements.

    “The other one is the exponential increase of computing resource requirements caused by blindly increase of the network size.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given
    § I. INTRODUCTION · p. 2

    Pooling layers filter the interference noise brought by unconscious jitter of the human body and decrease misjudgment.

    “Pooling layers are applied to lter the interference noise brought by the unconscious jitter of human body, in order to decrease the misjudgment;”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given
    § I. INTRODUCTION · p. 2

    Above mentioned deep convolutional neural networks ignored temporal dependencies on features and were not suitable to recognize real-time sensor signals.

    “Above mentioned deep convolutional neural networks ignored the temporal dependencies on the features, and was not suitable to recognize real-time sensor signals.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning
    § C. GOOGLENET AND INCEPTION MODULE · p. 4

    Larger network sizes often mean more parameters, making the expanded network more prone to overfit, especially when labeled training samples are limited.

    “One is that larger network sizes often mean more parameters, which makes the expanded network more prone to over t, especially when the labeled samples in the training set are limited.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § C. GOOGLENET AND INCEPTION MODULE · p. 4

    Extensive application of the pooling layer is useful to achieve dimensionality reduction, key feature extraction, and filtering.

    “Extensive application of the pooling layer is useful to achieve dimensionality reduction, key feature extraction, and ltering.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [9]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § I. INTRODUCTION · p. 2

    Most existing deep learning studies for HAR have used deep learning as a black box and scratched the data.

    “Most of these existing research have only used deep learning as a black box, and the data has been scratched. For example, Ronao and Cho [22] and Yang et al. [25] perform feature extraction of the sliding window using only shallow convolutional neural networks.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [22]

      Not checked in the free run.

    • cites [25] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § C. GOOGLENET AND INCEPTION MODULE · p. 4

    A 1x1 convolution network implements dimensionality reduction of data and linear combination of multi-channel data.

    “For example, a 1 1 convolution network implements dimensionality reduction of data and linear combination of multi-channel data.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [9]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given
    § B. LONG SHORT-TERM MEMORY · p. 3

    Ignoring temporal context beyond frame boundaries during modeling may limit recognition performance for more challenging tasks.

    “However, ignoring the temporal context beyond frame boundaries during modeling may limit the recognition performance for more challenging tasks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation
    § I. INTRODUCTION · p. 2

    Feature extraction methods such as symbolic representation, statistics of raw data, and transform coding are heuristic and require expert knowledge to design features.

    “The feature extraction methods such as symbolic representation [22], statistics of raw data [23] and transform coding [24] are widely applied in human activity recognition, but they are heuristic and require expert knowledge to design features [14].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [22]

      Not checked in the free run.

    • cites [23] source

      Not checked in the free run.

    • cites [24] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § I. INTRODUCTION · p. 2

    Chavarriaga et al.'s Naive Bayes and K-Nearest Neighbor classifier relied on hand-crafted features and could not find discriminative features to accurately distinguish different activities.

    “Chavarriaga et al. [21] used a Na ve Bayes(NB) and a K-Nearest Neighbor(KNN) classi er to recognize seven motions, such as walking, running and jumping. However, they relied on hand-crafted features and could not nd discriminative features to accurately distinguish different activities.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 7

    The Opportunity, PAMAP2, and Smartphone datasets are three of the most widely used public datasets for human activity recognition.

    “Furthermore, we generalize and verify these three models to three most widely used public datasets, namely Opportunity dataset [21], PAMAP2 dataset [37] and smartphone dataset [38].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [21] source

      Not checked in the free run.

    • cites [37]

      Not checked in the free run.

    • cites [38] [38] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, ''A public domain dataset for human activity recognition using smartphones,'' in Proc. ESANN , 2013, pp. 1 6. <!-- image --> CHENG XU received the B.E. and M.S. degrees from the University of Science and Technology Beijing, China, in 2012 and 2015, respectively, where he is currently pursuing the Ph.D. degree with the Data and Cyber-Physical System Lab. His research interests include patter recognition and the Internet of Things. He is a Student Member of the IEEE and CCF. <!-- image --> <!-- image --> DUO CHAI received the d

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § B. COMPARED WITH IMPROVED STATE-OF-THE-ARTS · p. 8

    In actual training, a large part of time is occupied in training LSTM, but the income is relatively little.

    “Therefore, in actual training process, a large part of time is often occupied in trainging process, but the income is relatively little.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning
    § D. PERFORMANCE MEASURE · p. 6

    F-measure considers the correct classification of each class as equally important and can evaluate the model better than precision.

    “F-measure ( F 1) considers the correct classi cation of each class as equally important. It takes into account both the precision and the recall of each class to compute the score and can evaluate the model better than the precision.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning
    § A. COMPARED WITH BASELINES AND STATE-OF-THE-ARTS · p. 8

    InnoHAR's superior performance could be due to the multi-scaled and spatial-temporal feature extraction characteristics from the combination of inception and GRU.

    “It could be because of the multi-scaled and spatial-temporal feature extraction characteristics, due to the combination of inception and GRU.”
    Quote in the paperSubstantial claim

  • May combine two claims. The quote only partly states this claim.

    Interpretationown data
    § C. EFFICIENCY ANALYSIS · p. 8

    InnoHAR can easily predict real-time activity on the MinnowBoard because the original test data is recorded at 30 Hz and the prediction speed is 65.09 pieces/s.

    “The original test data is recorded at 30 Hz, and our model can easily predict the real-time activity on this platform.”
    Quote in the paperSubstantial claim
29 of 29

  • The paper relies on this without support.

    Traditional feature engineering methods are becoming incapable in the face of extremely fast-growing waveform data with no obvious laws.

    “However, in the face of extremely fast-growing waveform data with no obvious laws, the traditional feature engineering methods are becoming more and more incapable.”

  • The paper relies on this without support.

    GRU performs better than LSTM in terms of time efficiency.

    “while GRU performs better in terms of time ef ciency.”

  • The paper relies on this without support.

    GRU and LSTM show better performance in dealing with long sequences based problems.

    “while in many existing experiments, GRU and LSTM show better performance in dealing with long sequences based problems”

  • The paper relies on this without support.

    Applying features extracted from the original signal to the neural network tends to improve performance.

    “The input of neural network is generally the original signal, however, applying features extracted from the original signal to the neural network tends to improve performance.”

  • The paper relies on this without support.

    The three public HAR datasets used are the most widely used public HAR datasets.

    “Through experimental veri cation on three most widely used public HAR datasets”

  • The paper relies on this without support.

    The MinnowBoard Turbot Dual Core Board is representative of embedded systems used in real-time human activity recognition applications.

    “The embedded system we used in this experiment was the MinnowBoard Turbot Dual Core Board [27]. It is a small embedded platform released by Intel that is equipped with Intel Atom E3826 processor, with 1.46 GHz clock speed and 2GB DDR3L 1067MT/s DRAM. It is an outstanding embedded platform with both performance and price advantages.”

  • The paper relies on this without support.

    The Opportunity, PAMAP2, and Smartphone datasets are representative of real-world complex human activity recognition, so performance on them generalizes to real-world HAR.

    “Furthermore, we generalize and verify these three models to three most widely used public datasets, namely Opportunity dataset [21], PAMAP2 dataset [37] and smartphone dataset [38].”
7 of 7

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?

  • In [227], GoogLeNet's Inception module combined with a GRU layer was used to build a HAR model.

    “In [227], the authors have used GoogLeNet’s Inception module combined with a GRU layer to build a HAR model.”

    From Deep Learning in Human Activity Recognition with Wearable Sensors: A Review on Advances (2022) · cites this paper as [227]

  • The model in [227] showed performance improvements on Opportunity, PAMAP2, and Smartphones datasets.

    “The proposed model was showed performance improvements on three public datasets (Opportunity, PAMAP2 and Smartphones datasets).”

    From Deep Learning in Human Activity Recognition with Wearable Sensors: A Review on Advances (2022) · cites this paper as [227]

  • Compared with the CNN model of Yang et al., LSTM-CNN has a significant increase of about 7% for the OPPORTUNITY dataset and is superior to the DeepConvLSTM model.

    “Compared with the CNN model of Yang et al., LSTM-CNN has a significant increase of about 7% for the OPPORTUNITY dataset and is superior to the DeepConvLSTM model.”

    From LSTM-CNN Architecture for Human Activity Recognition (2020) · cites this paper as [25]

  • InnoHAR demonstrated significantly higher performance and strong generalisation performance than state-of-the-art and baselines on a sliding window dataset.

    “On a sliding window dataset, InnoHAR demonstrated significantly higher performance and strong generalisation performance than state-of-the-art and baselines.”

    From Review of Wearable Devices and Data Collection Considerations for Connected Health (2021) · cites this paper as [170]

  • InnoHAR was generalised and validated using three datasets: Opportunity, PAMAP2, and Smartphone, with F-scores of 0.946, 0.935 and 0.945, respectively.

    “This model has been generalised and validated using three datasets: Opportunity, PAMAP2, and Smartphone, with F-scores of 0.946, 0.935 and 0.945, respectively [170].”

    From Review of Wearable Devices and Data Collection Considerations for Connected Health (2021) · cites this paper as [170]

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.

  • Temporal layers described as both LSTM and GRU

    Minor

    The temporal-feature section first says the model selects two LSTM layers, then says GRU is the implementation of those layers.

  • Results section identified as Section IV instead of V

    Minor

    The organization paragraph places the results and analysis in Section IV, but the provided section heading places them in Section V.

  • PAMAP2 windows described as non-overlapping and overlapping

    Minor

    The same windowing procedure is called non-overlapping and assigned 78% overlap.

  • Two different Opportunity test-window counts

    Minor

    The preprocessing section reports 9,984 test windows, while the efficiency analysis says the model predicted all 9,894 test windows.

4 of 4

Cites

Flagged first, then by citations

2 of the 38 references couldn’t be matched to a known paper.

Research works printed but not linked: 4

  1. 2019, Minnowboardorg
  2. gers, 2003, Learning precise timing with LSTM recurrent networks, J Mach Learn Res, 3, 115
  3. hammerla, 2016, Deep convolutional and recurrent models for human activity recognition using wearables
  4. yang, 2015, Deep convolutional neural networks on multichannel time series for human activity recognition, Proc Int Conf Artif Intell, 3995

Cited by

Sorted by citations

Related

Authors
Xu, C., Chai, D., He, J., Zhang, X., & Duan, S.
Affiliations
China Building Materials Academy (CN)
ORCID
Venue
IEEE Access
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2019
Volume
7
Pages
9893-9902
ISSN
2169-3536 (ISSN-L)
DOI
10.1109/access.2018.2890675
OpenAlex
W2907255073
Cited by
333
References
38
Citations by year
2026: 18 · 2025: 39 · 2024: 62 · 2023: 71 · 2022: 53 · 2021: 53 · 2020: 30 · 2019: 7
FWCI
16.01
Type
journal article
Language
en
Open access
Licence
ieeexplore.ieee.org/Xplorehelp/downloads/license-information/OAPA.html (vor)
Locations
  • IEEE Access (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Basic Research Program of China (2016YFC0901303)
  • National Natural Science Foundation of China (61671056, 61302065, 61304257, 61402033)
  • Natural Science Foundation of Beijing Municipality (4152036)
  • Tianjin Special Program for Science and Technology (16ZXCXSF00150)
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, sensor networks, body area networks, Inception module, recurrent neural networks, GRU, end-to-end learning, time series classification, feature extraction

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

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