CoChat
Sign Up
Cited by 863Context-Aware Activity Recognition SystemsHuman Pose and Action RecognitionGait Recognition and Analysis

LSTM-CNN Architecture for Human Activity Recognition

Open access
Xia, Kun, Huang, Jianguang, Wang, Hanyu·2020·IEEE Access
DOI 10.1109/access.2020.2982225
Open this paper in CoChat — free accountOpen PDF

Abstract

In the past years, traditional pattern recognition methods have made great progress. However, these methods rely heavily on manual feature extraction, which may hinder the generalization model performance. With the increasing popularity and success of deep learning methods, using these techniques to recognize human actions in mobile and wearable computing scenarios has attracted widespread attention. In this paper, a deep neural network that combines convolutional layers with long short-term memory (LSTM) was proposed. This model could extract activity features automatically and classify them with a few model parameters. LSTM is a variant of the recurrent neural network (RNN), which is more suitable for processing temporal sequences. In the proposed architecture, the raw data collected by mobile sensors was fed into a two-layer LSTM followed by convolutional layers. In addition, a global average pooling layer (GAP) was applied to replace the fully connected layer after convolution for reducing model parameters. Moreover, a batch normalization layer (BN) was added after the GAP layer to speed up the convergence, and obvious results were achieved. The model performance was evaluated on three public datasets (UCI, WISDM, and OPPORTUNITY). Finally, the overall accuracy of the model in the UCI-HAR dataset is 95.78%, in the WISDM dataset is 95.85%, and in the OPPORTUNITY dataset is 92.63%. The results show that the proposed model has higher robustness and better activity detection capability than some of the reported results. It can not only adaptively extract activity features, but also has fewer parameters and higher accuracy.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes an LSTM-CNN architecture that feeds raw mobile sensor data into a two-layer LSTM followed by convolutional layers, with global average pooling and batch normalization to reduce parameters and speed convergence. It reports high accuracy/F1 on UCI-HAR, WISDM, and OPPORTUNITY and compares favorably with CNN and DeepConvLSTM baselines.

Key findings

  • On three public datasets, reported overall accuracy/F1 is UCI-HAR 95.78% (abstract/conclusion; experimental section says 95.80% accuracy), WISDM 95.85% (abstract/conclusion; experimental section says 95.75% accuracy), and OPPORTUNITY 92.63%; without the Null class, OPPORTUNITY accuracy is 87.58%.
  • Compared with the CNN of Yang et al. and DeepConvLSTM under the same scenario, LSTM-CNN increased about 7% on OPPORTUNITY and averaged 3% on UCI-HAR and WISDM by F1 score.
  • Replacing the fully connected layer with GAP reduced parameters: model B has 27,462 parameters versus more than 502 thousand for model A, about 94% less, with almost the same performance; adding BN raised F1 to 93.35%, and adding LSTM layers raised UCI-HAR F1 to 95.78%.
  • Hyperparameter tests on UCI-HAR found Adam was the best optimizer, batch size 192 gave the highest accuracy, and 192 filters in the second convolutional layer gave F1 96.38% (4% over 64 filters) but increased model parameters by more than 70%.

Methods

  • Architecture: preprocessed sensor data passes through two LSTM layers (32 memory cells each, 64 total), then two convolutional layers (64 filters of size 1×5 with stride 2; 128 filters of size 1×3 with stride 1), with max-pooling between them, followed by GAP, BN, and a dense Softmax output layer.
  • Data: UCI-HAR (30 subjects, 6 activities, 50 Hz, 748,406 samples), WISDM (36 subjects, 6 activities, 20 Hz, 1,098,209 samples, unbalanced), and OPPORTUNITY (4 subjects, 17 activities plus Null class, 30 Hz, 113 channels). Preprocessing used linear interpolation, 0–1 normalization, and sliding windows of length 128 for UCI/WISDM and 24 for OPPORTUNITY, with 50% overlap.
  • Training/evaluation: Keras with TensorFlow backend, cross-entropy loss, Adam optimizer, batch size 192, 200 epochs, learning rate 0.001; evaluated with accuracy and F1 score, confusion matrices, and comparisons to CNN and DeepConvLSTM.

Limitations

  • The paper states WISDM and OPPORTUNITY are imbalanced, so overall accuracy is not an appropriate measure and F1 is used; it also notes relatively poor discrimination between sitting and standing in UCI-HAR, with recall and precision in the range of 92%–93%.
  • The paper states LSTM layers slow computation because each time step depends on the previous one and cannot be computed in parallel; LSTM-CNN took 9,416 ms per epoch versus 5,156 ms for model D, and GAP can shift training pressure to convolutional layers and slow convergence, which BN was added to address.
  • Evident from the text: reported metrics are inconsistent across sections (e.g., UCI-HAR 95.78% vs 95.80%; WISDM 95.85% vs 95.75%), and OPPORTUNITY evaluation uses only four subjects, limiting generalization claims.

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

Integrity signals

Registry

Crossref · 10.1109/access.2020.2982225

Cites

39 of 39 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 10 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The LSTM-CNN model achieved an overall accuracy of 95.78% on the UCI-HAR dataset.

    “Finally, the overall accuracy of the model in the UCI-HAR dataset is 95.78%, in the WISDM dataset is 95.85%, and in the OPPORTUNITY dataset is 92.63%.”
    Quote in the paperSubstantial claim
  • Findingown data

    The LSTM-CNN model achieved an overall accuracy of 92.63% on the OPPORTUNITY dataset.

    “Finally, the overall accuracy of the model in the UCI-HAR dataset is 95.78%, in the WISDM dataset is 95.85%, and in the OPPORTUNITY dataset is 92.63%.”
    Quote in the paperSubstantial claim
  • Findingown data

    The LSTM-CNN model achieved an overall accuracy of 95.85% on the WISDM dataset.

    “Finally, the overall accuracy of the model in the UCI-HAR dataset is 95.78%, in the WISDM dataset is 95.85%, and in the OPPORTUNITY dataset is 92.63%.”
    Quote in the paperSubstantial claim
  • Findingown data

    Model parameters increase by more than 70% when increasing the number of filters.

    “However, the model parameters increase by more than 70%.”
    Quote in the paperNot a substantial claim
  • Findingown data

    On the OPPORTUNITY dataset, an overall accuracy of 92.63% was achieved.

    “Finally, an overall accuracy of 92.63% was achieved.”
    Quote in the paperSubstantial claim
  • May combine two claims. 28
  • Classified as probably not a substantial claim. 10
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    Model A has more than 502 thousand model parameters, with the parameters of fully-connected layers occupying the most.

    “Accordingly, there are more than 502 thousand model parameters, of which the parameters of fully-connected layers occupy the most.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    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.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [38] source

      Not checked in the free run.

    • cites [25] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    F1 score reaches 96.38% when the number of filters is 192, outperforming 64 filters by 4%.

    “F 1 score reaches 96.38% when the number of filters is selected as 192, which outperforms when the number of filters is 64 by 4%.”
    Quote in the paperSubstantial claim

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

    Findingown data

    Model A takes 1681 milliseconds per epoch in the training stage.

    “And it takes 1681 milliseconds per epoch in the training stage.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Model B has only 27462 parameters, about 94% less than model A, while performance remains almost the same.

    “The parameters of the model B are only 27462, which is about 94% less than that of the model A while the performance remains almost the same.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    The model parameters were greatly reduced by adding the GAP layer to the network.

    “It should be noted that the model parameters have been greatly reduced under adding the GAP layer to the network.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The computation speed of model D and LSTM-CNN is 5156 milliseconds per epoch and 9416 milliseconds per epoch, respectively.

    “They are 5156 milliseconds per epoch and 9416 milliseconds per epoch, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The model trained by Adam optimizer has the best fitting effect and the most stable fluctuation of gradient descent curve.

    “It can be seen that the model trained by Adam optimizer has the best fitting effect and the fluctuation of gradient descent curve is the most stable.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The final F1 score reached 95.78%, 95.85% and 92.63% on the UCI-HAR, WISDM and OPPORTUNITY datasets, respectively.

    “Eventually, the F 1 score reached 95.78%, 95.85% and 92.63% on the UCI-HAR, WISDM and OPPORTUNITY datasets, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    LSTM-CNN outperforms the other two models on the UCI-HAR and WISDM datasets, with the best-reported result increasing by an average of 3%.

    “It can also be seen that LSTM-CNN outperforms the other two models on the UCI-HAR and WISDM datasets, with the best-reported result increasing by an average of 3%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On the UCI-HAR dataset, 2940 instances were correctly classified and the overall accuracy reached 95.80%.

    “For the UCI-HAR dataset, there were 2940 instances that have been correctly classified, and the overall accuracy reached 95.80%.”
    Quote in the paperSubstantial claim

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

    Interpretationreasoning

    The poor discrimination between sitting and standing may be because the two activities are similar from the perspective of motion sensors and it is difficult to mine deeper information only by acceleration and angular velocity data.

    “The main reason may be that the two activities are similar from the perspective of motion sensors.It is difficult to mine deeper information only by acceleration and angular velocity data.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Deep learning methods could automatically extract appropriate features from raw sensor data during the training phase and present low-level original temporal features with high-level abstract sequences.

    “To address this problem, researchers have turned to deep learning methods that could automatically extract appropriate features from raw sensor data during the training phase and present the low-level original temporal features with high-level abstract sequences.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [24]- [27]

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    Existing deep learning models for human activity recognition have relatively complex network structures and large numbers of parameters, resulting in high computational cost and difficulty for real-time performance.

    “Although the above models could generally recognize human activities, the overall network structure is relatively complex.In addition, these models have a large number of parameters, which results in high computational cost.It is difficult to be used in occasions that require high real-time performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Compared with methods proposed in other literature, the LSTM-CNN model shows consistent superior performance and good generalization.

    “To sum up, compared with the methods proposed in other literatures, the LSTM-CNN model shows consistent superior performance and has good generalization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The proposed model achieves high recognition accuracy while greatly simplifying the model structure.

    “The model proposed in this paper not only could achieve high recognition accuracy but also greatly simplify the model structure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Adding a BN layer after the GAP layer stabilizes the output of the upper layer, speeds up convergence, and improves accuracy.

    “The model C adds a BN layer after the GAP layer to stabilize the output of the upper layer.It speeds up the convergence of the model and improves accuracy.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    GAP sums out the spatial information, so it is more robust to the spatial transformation of the input.

    “Furthermore, GAP sums out the spatial information, so it is more robust to the spatial transformation of the input.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    The GAP layer has no parameter to optimize, thus it reduces global model parameters.

    “There is no parameter to optimize in the GAP layer.Thus, it achieves the goal of reducing global model parameters.”
    Quote in the paperSubstantial claim

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

    Study design and conductown data

    For the UCI-HAR dataset, recordings of 22 subjects were used to build the training set and the remaining subjects were used to build the test set.

    “Among them, the recordings of 22 subjects were used to build the training set, and the rest was used to build the test set.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductnone given

    For the WISDM dataset, the training set was composed of recordings of 30 subjects and the remaining 6 subjects' recordings were used to build the test set, with the two parts completely separate.

    “The training set is composed of the recordings of 30 subjects and the remaining recordings of 6 subjects are used to build the test set.The two parts are completely separate.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Interpretationown data

    The LSTM-CNN model has higher accuracy than some reported results.

    “It can not only adaptively extract activity features, but also has fewer parameters and higher accuracy.”
    • Also stated in: § II. RELATED WORK · plain
      “The results show that the proposed model not only has high accuracy but also has good generalization ability and fast convergence speed.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Agarwal et al.'s model was only evaluated on one dataset with six activities, which does not prove that it has good generalization ability.

    “Although the proposed model has high accuracy and brief architecture, it was only evaluated on one dataset which has just six activities, which does not prove that the proposed model has good generalization ability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Machine learning methods may rely heavily on heuristic manual feature extraction in most daily human activity recognition tasks, which is limited by human domain knowledge.

    “Machine learning methods may rely heavily on heuristic manual feature extraction in most daily human activity recognition tasks.It is usually limited by human domain knowledge [23].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23] source

      Not checked in the free run.

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

    Premisereasoning

    The BN layer normalizes and reconstructs the input data on each batch of training samples to ensure the stability of the output of the previous layer, so as to improve the training speed and accuracy.

    “The BN layer normalizes and reconstructs the input data on each batch of training samples to ensure the stability of the output of the previous layer, so as to improve the training speed and accuracy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    During training, the distribution of input data of each layer continuously changes due to weight updates, leading to difficulty in network training and slowing down convergence.

    “During the training process, the distribution of input data of each layer would continuously change due to the weight parameters of the upper layer are constantly updated.Therefore, it is necessary to change the weight parameters to adapt to this new distribution, which leads to difficulty in network training and slows down the convergence speed.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    CNN has gained increasing popularity because of its ability to learn unique representations from images or speech.

    “CNN has gained increasing popularity because of its ability to learn unique representations from images or speech [34].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [34] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Using the global average pooling layer and batch normalization layer to replace the fully-connected layer is effective, and using LSTM to extract temporal information improves model performance.

    “To sums up, the strategy of using the global average pooling layer and batch normalization layer to replace the fully-connected layer is effective.Moreover, the method of using LSTM to extract the temporal information to improve model performance is favorable.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Recordings of activities based on mobile sensors are temporal sequences.
    • Rests on an assumption: LSTM has the ability to capture temporal information from sequential data.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisereasoning

    The fully-connected layer at the end of the model merges features extracted from the upper layer, making up for the shortcomings of the GAP layer.

    “Each node of the fully-connected layer is connected to the nodes of the upper layer so that the features extracted from the upper layer could be merged.It makes up for the shortcomings of the GAP layer in this regard.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    The InnoHAR model [29] achieved excellent results but converged hardly, causing a lot of time to be wasted in the training stage.

    “The results showed an excellent effect, but the model converged hardly, causing a lot of time to be wasted in the training stage.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [29] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    In the model Krizhevsky [35], the first fully-connected layer FC1 has 4096 nodes, and the output of the upper pooling layer Max-Pool3 has 9216 nodes, resulting in more than 37 million weight parameters between them, which consumes a lot of memory and computational cost.

    “For instance, in the model Krizhevsky [35], the first fully-connected layer FC1 has 4096 nodes, and the output of the upper pooling layer Max-Pool3 has 9216 nodes.Thus, there would be more than 37 million weight parameters between the MaxPool3 layer and the FC1 layer, which would consume a lot of memory and computational cost.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [35]

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data

    The model avoids complex feature extraction and has high recognition accuracy with few model parameters.

    “It can not only avoid complex feature extraction but also has high recognition accuracy under the premise of a few model parameters.”
    Quote in the paperSubstantial claim
32 of 32

  • The paper relies on this without support.

    LSTM is a variant of RNN that is more suitable for processing temporal sequences.

    “LSTM is a variant of the recurrent neural network (RNN), which is more suitable for processing temporal sequences.”

  • The paper relies on this without support.

    LSTM can eliminate the gradient vanishing limitation of RNN.

    “LSTM is a variety of RNN, which could eliminate this limitation.”

  • The paper relies on this without support.

    LSTM has great advantages in feature extraction of sequence data than convolutional neural networks due to its special memory cells.

    “LSTM has great advantages in feature extraction of sequence data than convolutional neural networks due to its special memory cells.”

  • The paper relies on this without support.

    RNN can take advantage of the chronological relationship between sensor readings.

    “RNN could take advantage of the chronological relationship between sensor readings.”

  • The paper relies on this without support.

    RNN has the problem of gradient vanishing, which hinders the ability of the network to model between raw sensor data and human activities in a long context window.

    “Although RNN has the ability to capture temporal information from sequential data, it has the problem of gradient vanishing, which hinders the ability of the network to model between raw sensor data and human activities in a long context window.”

  • The paper relies on this without support.

    The more convolution kernels are used, the more hidden features could be mined in the input samples.

    “Generally speaking, the more convolution kernels are used, the more hidden features could be mined in the input samples.”

  • The paper relies on this without support.

    The max-pooling layer filters interference noise caused by the unconscious jitter of the human body.

    “It serves two purposes.One is to reduce the parameters while maintaining dominant features, the other is to filter the interference noise caused by the unconscious jitter of the human body.”

  • The paper relies on this without support.

    The WISDM and OPPORTUNITY datasets are imbalanced.

    “The WISDM and OPPORTU-NITY mentioned above are both imbalanced datasets.”

  • The paper relies on this without support.

    Hyper-parameters have a great impact on classification model performance.

    “Hyper-parameters have a great impact on the classification model performance.”

  • The paper relies on this without support.

    The weight parameters of CNN mainly concentrate on the fully-connected layer.

    “The weight parameters of CNN mainly concentrate on the fully-connected layer.”

  • The paper relies on this without support.

    Recordings of activities based on mobile sensors are temporal sequences.

    “The recordings of activities based on mobile sensors are temporal sequences and LSTM has the ability to capture temporal information from sequential data.”

  • The paper relies on this without support.

    LSTM has the ability to capture temporal information from sequential data.

    “The recordings of activities based on mobile sensors are temporal sequences and LSTM has the ability to capture temporal information from sequential data.”

  • The paper relies on this without support.

    When training LSTM layers, the calculation of each time step depends on the output of the previous time step.

    “When training LSTM layers, the calculation of each time step depends on the output of the previous time step.”
13 of 13

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?

  • The more convolution kernels are used, the more hidden features of the input samples can be extracted.

    “In general, it is examined that the more convolution kernels are used, the more hidden features of the input samples could be extracted [48].”

    From Deep CNN-LSTM With Self-Attention Model for Human Activity Recognition Using Wearable Sensor (2022) · cites this paper as [48]

  • Our chosen model achieved the best testing result of 99.93% for H-Activity dataset and 98.76% and 93.11% for MHEALTH and UCI-HAR dataset respectively.

    “In contrast to these three models, i.e.,sequential LSTM layers (LSTM-CNN) before Convolution layer [48], sequential Convolution and dropout layers (CNN-LSTM) before LSTM layer [59], parallel LSTM layers (Parrallel LSTM-CNN) with Convolution layer [60], our chosen model achieved the best testing result of 99.93% for H-Activity dataset and 98.76% and 93.11% for MHEALTH and UCI-HAR dataset respective”

    From Deep CNN-LSTM With Self-Attention Model for Human Activity Recognition Using Wearable Sensor (2022) · cites this paper as [48]

  • All of the above-mentioned models [48], [59], [60] were implemented with exact parameters described in the respective literature except the input and output layer to match our dataset.

    “Hence, all of the above-mentioned models [48], [59], [60] were implemented with exact parameters described in the respective literature except the input and output layer to match our dataset.”

    From Deep CNN-LSTM With Self-Attention Model for Human Activity Recognition Using Wearable Sensor (2022) · cites this paper as [48]

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.

  • UCI-HAR overall accuracy has two reported values

    Major

    The abstract reports 95.78% overall accuracy for UCI-HAR, whereas the results section reports 95.80% for the same measure.

  • WISDM waveform duration conflicts with its sampling rate

    Minor

    At the stated WISDM sampling rate of 20 Hz, 128 points span 6.4 seconds, not 2.56 seconds.

    Reported 2.56 · recomputed 6.4

  • UCI-HAR is incorrectly described as having the most volunteers

    Minor

    The paper says UCI-HAR has the largest number of volunteers at 30, but states that WISDM has 36 subjects.

  • Adding an LSTM layer to model D leaves the stated total unchanged

    Minor

    Model D is described as having two LSTM layers. The paper then says it added another layer to model D for a total of two layers.

  • WISDM overall accuracy has two reported values

    Minor

    The abstract reports 95.85% overall accuracy for WISDM, whereas the results section reports 95.75%.

5 of 5

Cites

Flagged first, then by citations

1 of the 40 references couldn’t be matched to a known paper.

Cited by

Sorted by citations

Related

Authors
Xia, K., Huang, J., & Wang, H.
Affiliations
University of Shanghai for Science and Technology (CN)
ORCID
Venue
IEEE Access
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2020
Volume
8
Pages
56855-56866
ISSN
2169-3536 (ISSN-L)
DOI
10.1109/access.2020.2982225
OpenAlex
W3011785450
Cited by
851
References
40
Citations by year
2026: 68 · 2025: 169 · 2024: 169 · 2023: 185 · 2022: 136 · 2021: 83 · 2020: 14 · 2017: 2 · 2016: 2 · 2014: 3 · 2013: 1 · 2012: 1
FWCI
40.64
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Access (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
National Natural Science Foundation of China (512070901, 51637001)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Human Pose and Action Recognition
  • Physical Sciences › Engineering › Biomedical Engineering › Gait Recognition and Analysis
Keywords
human activity recognition, long short-term memory, convolutional neural networks, mobile sensor data, wearable computing, automatic feature extraction, global average pooling, batch normalization, UCI HAR dataset, WISDM dataset, Opportunity dataset

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