Accuracy Comparison of CNN, LSTM, and Transformer for Activity Recognition Using IMU and Visual Markers
Open accessThe DOI matches its Crossref record; 47 of the 47 references it cites were found in a registry or an index.
Abstract
Human activity recognition (HAR) has applications ranging from security to healthcare. Typically these systems are composed of data acquisition and activity recognition models. In this work, we compared the accuracy of two acquisition systems: Inertial Measurement Units (IMUs) vs Movement Analysis Systems (MAS). We trained models to recognize arm exercises using state-of-the-art deep learning architectures and compared their accuracy. MAS uses a camera array and reflective markers. IMU uses accelerometers, gyroscopes, and magnetometers. Sensors of both systems were attached to different locations of the upper limb. We captured and annotated 3 datasets, each one using both systems simultaneously. For activity recognition, we trained 8 architectures, each one with different operations and layers configurations. The best architectures were a combination of CNN, LSTM, and Transformer achieving test accuracy from 89% to 99% on average. We evaluated how feature selection reduced the sensors required. We found IMU and MAS data were able to distinguish correctly the arm exercises. CNN layers at the beginning produced better accuracy on challenging datasets. IMU had advantages over other aquisition systems for activity recognition. We analyzed the relations between models accuracy, signal waveforms, signals correlation, sampling rate, exercise duration, and window size. Finally, we proposed the use of a single IMU located at the wrist and a variable-size window extraction.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Figures


![FIGURE 3 Exercises (a) flexion-extension [37], (b) pronation-supination.](https://figures.cochat.ai/paper-figures/cde038770e744d9e8984b1fda96bb6a1/accuracy-comparison-of-cnn-lstm-and-transformer-for-activity-recognition-using-figure-3.png)
![FIGURE 4 Shoulder Abduction-Adduction. (a) neutral position, (b) abduction from 0 ◦ to 60 ◦ [37]](https://figures.cochat.ai/paper-figures/a119b705b3dd4211a9cdf42fef9a33d3/accuracy-comparison-of-cnn-lstm-and-transformer-for-activity-recognition-using-figure-4.png)
![FIGURE 6 Vertical Shoulder Flexion-extension. (a) low amplitude (45 ◦ - 50 ◦ ), (b) high amplitude (180 ◦ ) [37]](https://figures.cochat.ai/paper-figures/064574e746994585a18deacdaf5af0b9/accuracy-comparison-of-cnn-lstm-and-transformer-for-activity-recognition-using-figure-6.png)
![FIGURE 7 Internal and external rotation of the shoulder. (a) internal rotation of 30 ◦ , (b) external rotation of 80 ◦ [37]](https://figures.cochat.ai/paper-figures/65854d68c6b14562831406e847b43204/accuracy-comparison-of-cnn-lstm-and-transformer-for-activity-recognition-using-figure-7.png)















Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
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 check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“For food intake recognition [24], it has been proven possible to only annotate and recognize each cycle of a single gesture (micro movements) based on seq-to-seq classification.”
From DS-MS-TCN: Otago Exercises Recognition With a Dual-Scale Multi-Stage Temporal Convolutional Network (2024) · cites this paper as [24]
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Breiman, L. (2001) · Machine Learning · doi:10.1023/a:1010933404324 · Cited by 133,538 · cited 1×
“The algorithm used for feature selection, random forest (RF) [50], [51], is an ensemble learning method combining multiple decision trees to improve the accuracy of the model.”
MethodVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 2×
“This means that all parts of the sequence are treated by their real importance, and the overall context is considered, prioritizing words with higher weight, allowing the model to focus on the right element of the input to predict the next element of the output [43], [44].”
BackgroundHochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 1×
“It comprises memory cells and gates that allow the network to store or discard information over time selectively [39].”
BackgroundExploiting Generative AI to Scale up Intelligent Tutoring Systems
Jakubův, J., Chvalovský, K., Goertzel, Z., Kaliszyk, C., Olšák, M., Piotrowski, B., Schulz, S., Suda, M., & Urban, J. (2023) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.itp.2023.19 · Cited by 76,658
Transformers: State-of-the-Art Natural Language Processing
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations · doi:10.18653/v1/2020.emnlp-demos.6 · Cited by 8,741
Greff, K., Srivastava, R. K., Koutnik, J., Steunebrink, B. R., & Schmidhuber, J. (2017) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2016.2582924 · Cited by 6,955 · cited 1×
“Additionally, each gate has its own set of learnable parameters, allowing the network to adaptively adjust the amount of information stored or discarded based on the input data, making LSTM effective in processing sequential data [40].”
BackgroundHo, T. K. (2002) · Proceedings of 3rd International Conference on Document Analysis and Recognition · doi:10.1109/icdar.1995.598994 · Cited by 5,054
Introduction to modern photogrammetry
Mikhail, E. M., Bethel, J., & McGlone, J. C. (2001) · Cited by 1,019
Yadav, S., & Shukla, S. (2016) · 2016 IEEE 6th International Conference on Advanced Computing (IACC) · doi:10.1109/iacc.2016.25 · Cited by 773 · cited 1×
“Additionally, in every training process, the dataset was split into 5 parts using k-fold cross validation [38] to alleviate the effects of small datasets and class imbalance.”
MethodA Comparative Study on Transformer vs RNN in Speech Applications
Karita, S., Chen, N., Hayashi, T., Hori, T., Inaguma, H., Jiang, Z., Someki, M., Soplin, N. E. Y., Yamamoto, R., Wang, X., et al. (2019) · 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) · doi:10.1109/asru46091.2019.9003750 · Cited by 727 · cited 1×
“This means that all parts of the sequence are treated by their real importance, and the overall context is considered, prioritizing words with higher weight, allowing the model to focus on the right element of the input to predict the next element of the output [43], [44].”
Background
Research works printed but not linked: 7
- T. Kam Ho, ''Random decision forests,'' in Proc. 3rd Int. Conf. Document Anal. Recognit., 1995, pp. 278-282.
- A. Liaw and M. Wiener, ''Classification and regression by randomforest,'' R News, vol. 2, no. 3, pp. 18-22, 2002.
- P. Rathnayakas. (2021). ETFA-workshop. [Online]. Available: https://github.com/CDAC-lab/ETFA-Workshop
- I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge, MA, USA: MIT Press, 2016.
- R. Tokarczyk and T. Mazur, ''Photogrammetry-principles of operation and application in rehabilitation,'' Med. Rehabil., vol. 10, pp. 31-38, Jan. 2006.
- 2022, Documentation
- A. Kapandji, Fisiología Articular, Tomo I, Tomo II, Tomo III. Madrid, Spain: Panamericana, 2010.
Cited by
RNN-LSTM: From applications to modeling techniques and beyond—Systematic review
Al-Selwi, S. M., Hassan, M. F., Abdulkadir, S. J., Muneer, A., Sumiea, E. H., Alqushaibi, A., & Ragab, M. G. (2024) · Journal of King Saud University - Computer and Information Sciences · doi:10.1016/j.jksuci.2024.102068 · Cited by 505
Long Short-Term Memory Networks: A Comprehensive Survey
Krichen, M., & Mihoub, A. (2025) · AI · doi:10.3390/ai6090215 · Cited by 179
Comparison of LSTM and Transformer for Time Series Data Forecasting
Sonata, I., & Heryadi, Y. (2024) · International Conference on Informatics and Computational Sciences (ICICoS) · doi:10.1109/icicos62600.2024.10636892 · Cited by 31
“The use of Transformers was also carried out by Trujillo-Guerrero et al. [12] to recognize human activity based on Inertial Measurement Units (IMUs) and Movement Analysis Systems (MAS) which are also time series data.”
MethodOtago Exercises Monitoring for Older Adults by a Single IMU and Hierarchical Machine Learning Models
Shang, M., Dedeyne, L., Dupont, J., Vercauteren, L., Amini, N., Lapauw, L., Gielen, E., Verschueren, S., Varon, C., De Raedt, W., et al. (2024) · IEEE Transactions on Neural Systems and Rehabilitation Engineering · doi:10.1109/tnsre.2024.3355299 · Cited by 21
“Similar results were also reported by a previous study, reporting that CNN-LSTM outperformed CNN, LSTM, and Transformer for IMU-based activity recognition [18].”
MethodResultMachine learning models for wearable-based human activity recognition: A comparative study
Ciortuz, G., Pour, H. H., Irshad, M. T., Nisar, M. A., Huang, X., & Fudickar, S. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.130911 · Cited by 18
A Systematic Review on Custom Data Gloves
Belcamino, V., Carfì, A., & Mastrogiovanni, F. (2024) · IEEE Transactions on Human-Machine Systems · doi:10.1109/thms.2024.3394674 · Cited by 20
“D convolutional neural networks [175], transformers [176], generative adversarial networks [177], and variational autoen-coders [178].”
BackgroundTransformers model for DDoS attack detection: A survey
Farias, E. P., de Neira, A. B., Borges, L. F., & Nogueira, M. (2025) · Computer Networks · doi:10.1016/j.comnet.2025.111433 · Cited by 15
A Comprehensive Overview and Comparative Analysis of CNN, RNN-LSTM and TransformerPreprint
Ali, Z. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5175090 · Cited by 17
Yunardi, R. T., Sardjono, T. A., & Mardiyanto, R. (2024) · IEEE Access · doi:10.1109/access.2024.3451495 · Cited by 14
“The observation from recent studies showed the ability of deep learning models to improve classification performance through the architectures of neural networks processing the data related to the features [17].”
BackgroundHossain, M. B., LaMunion, S. R., Crouter, S. E., Melanson, E. L., & Sazonov, E. E. (2024) · Sensors · doi:10.3390/s24103046 · Cited by 9
Zhao, C., Jin, Y., An, R., Uchitomi, H., & Miyake, Y. (2025) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2025.111561 · Cited by 8
Kadam, S., & Kim, D. I. (2025) · IEEE Internet of Things Journal · doi:10.1109/jiot.2025.3581778 · Cited by 8
“While LSTMs excel in handling long-term dependencies, Transformer models, which take advantage of parallel processing, are better suited for large-scale tasks that require managing intricate long-term dependencies [25], [30].”
Residual network with self-adaptive time step size
Li, X., Zou, X., & Liu, W. (2024) · Pattern Recognition · doi:10.1016/j.patcog.2024.111008 · Cited by 7
Accurate power consumption estimation method makes walking robots energy efficient and quiet
Valsecchi, G., Vicari, A., Tischhauser, F., Garabini, M., & Hutter, M. (2024) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros58592.2024.10802798 · Cited by 11
“LSTMs have demonstrated superior performance in such cases, making them a natural choice for our predictive model [22].”
Moghaddam, E. S., Ellis, C., & Hejrati, B. (2025) · IEEE Access · doi:10.1109/access.2025.3606339 · Cited by 6
“Inertial measurement units (IMUs) are often the preferred type of body-worn sensors due to their ability to provide precise and consistent measurements of motion and orientation without limitations such as blind spots or environmental dependencies [11].”
Wei, R., Chen, Z., Wang, Q., Yongsheng, D., Wang, H., Jiang, F., Liu, D., & Wang, X. (2025) · Energies · doi:10.3390/en18143848 · Cited by 6
DS-MS-TCN: Otago Exercises Recognition With a Dual-Scale Multi-Stage Temporal Convolutional Network
Shang, M., Dedeyne, L., Dupont, J., Vercauteren, L., Amini, N., Lapauw, L., Gielen, E., Verschueren, S., Varon, C., De Raedt, W., et al. (2024) · IEEE Journal of Biomedical and Health Informatics · doi:10.1109/jbhi.2024.3455426 · Cited by 5
“However, transformers did not outperform CNN-LSTM according to the previous study for IMU-based activity recognition [24].”
Leveraging Swin Transformer for advanced sentiment analysis: a new paradigm
Rajput, G. K., Srivastava, S. K., & Gupta, N. (2025) · Cognitive Neurodynamics · doi:10.1007/s11571-025-10378-z · Cited by 6
Zi, Y., van de Ven, S. R. B., de Geus, E. J. C., & Chen, P. (2026) · Interactive Journal of Medical Research · doi:10.2196/76601 · Cited by 5
“Other reviews have focused on specific aspects, such as calibration techniques [45], sport-specific movements [46], or compared DL architectures like CNN and LSTM [47].”
Martínez-Pascual, D., Vales, Y., Batanero, R. M., Catalan, J. M., Blanco, A., & García-Aracil, N. (2026) · IEEE Transactions on Neural Systems and Rehabilitation Engineering · doi:10.1109/tnsre.2026.3669220 · Cited by 4
“…a Time LeNet model (t-LeNet) [45], a Time Convolutional Network (Time-CNN) [46], an LSTM model, an ENCODER model [39], and the model 2LConv from [47], which achieved the maximum performance when compared with other SOTA models for [17], RehabNet was also implemented in order to compare the…”
Chen, G., Takegawa, Y., Matsumura, K., Watanabe, H., & Hirata, K. (2025) · Sensors and Materials · doi:10.18494/sam5359 · Cited by 4
Wang, S., Hong, Z., Min, Q., Zou, D., Zhao, Y., Qi, R., & Zhao, T. (2025) · Energies · doi:10.3390/en18112934 · Cited by 4
Benachour, Y., Flitti, F., Maloukh, L., Far, A. B., Boutellaa, E., Bentoumi, M., Rai, M. C. E., Aburaed, N., Ali, K., Rehman, M. U., et al. (2026) · IEEE Access · doi:10.1109/access.2026.3707566 · Cited by 4
“Additional contributions arise from Wearable Computing and Sensor Systems (WC, 28), Neurorehabilitation and Neural Engineering (NE, 22), Human Movement Analysis and Biomechanics (HMA, 12), Biomedical Signal Processing and Computational Physiology (BSP, 9), Digital Health and e-Rehabilitation (DH,…”
InfluentialWolling, F., Kostolani, D., Trollmann, P., & Michahelles, F. (2025) · ACM International Symposium on Wearable Computers (ISWC) · doi:10.1145/3715071.3750419 · Cited by 3
“CNNs remain a popular choice due to their high capacity to extract distinctive features from raw data [41], and the integration with other architectures such as LSTMs [10] or Transformers [34, 53] achieves cutting-edge performance.”
Conforti, P. M., Nardelli, P., Fanti, A., & Russo, P. (2025) · IEEE Transactions on Artificial Intelligence · doi:10.1109/tai.2025.3611909 · Cited by 1
“A plethora of deep architectures, including recurrent neural networks (RNNs), convolutional neural networks (CNNs), and Transformers, are available in the literature for signal analysis and time series prediction, each with its advantages and disadvantages [2].”
Wearable interface for real-time gait phase recognition using sensor networks
Mohamed, S. A., & Martínez-Hernández, U. (2026) · Applied Soft Computing · doi:10.1016/j.asoc.2026.115005 · Cited by 1
Mekruksavanich, S., & Jitpattanakul, A. (2026) · Technologies · doi:10.3390/technologies14050272 · Cited by 1
A Comparative Study of Human Activity Recognition: Motion, Tactile, and Multi-modal Approaches
Belcamino, V., Le, N. M. D., Luu, Q. K., Carfì, A., Ho, V. A., & Mastrogiovanni, F. (2026) · Lecture notes in networks and systems · doi:10.1007/978-3-032-33874-7_11 · Cited by 1
Chawaphan, P., Maneetham, D., & Crisnapati, P. N. (2025) · International Journal of Innovative Research and Scientific Studies · doi:10.53894/ijirss.v8i3.6805 · Cited by 1
“This computational efficiency and ability to capture global dependencies make the Transformer ideal for handling complex tasks with long-range interactions [35].”
Multi-Scale Multi-Headed Attention Framework for Wearable IMU-based Exercise Recognition
Reddy, J. S., Sharma, S., Ratna, S., & Bakshi, T. (2025) · International Conference on Data Science and Information System (ICDSIS) · doi:10.1109/icdsis65355.2025.11071220
“[21] showed LSTMs outperform CNNs in capturing the temporal dynamics of exercise movements.”
Genomförbarhet av träningsigenkänning på Exxentric kBox med IMU-sensor
Yuping, H. (2024) · KTH Publication Database DiVA (KTH Royal Institute of Technology)
Shen, A., Chen, X., Xia, Q., Wang, Y., & Xu, H. (2026) · BioMedical Engineering OnLine · doi:10.1186/s12938-026-01513-w
“In other studies, arm motion pattern data captured by inertial measurement units (IMU) and motion analysis systems (MAS) were analyzed using CNN and transformer–LSTM algorithms by Trujillo-Guerrero [27], achieving recognition rates ranging from 89 to 94%.”
Fokam, C. R. K., & Yaddaden, Y. (2026) · International Conference on Electronics, Computers and Artificial Intelligence (ECAI) · doi:10.1109/ecai69016.2026.11613697
“Such transformations improve the representation of gesture dynamics and help highlight transitions between different phases of motion [4], [5].”
Islam, A., Rahman, A., Salgado, A. A. V., Scharf, R. J., Blemker, S. S., & Barnes, L. E. (2026) · IEEE International Conference on Healthcare Informatics (ICHI) · doi:10.1109/ichi69079.2026.00045
“Deep learning architectures such as CNNs, LSTMs, and Transformers have advanced IMU-based HAR, enabling end-to-end feature learning and improved robustness [19], [20].”
Kostolani, D., Wolling, F., Schlund, S., & Michahelles, F. (2026) · Frontiers in Computer Science · doi:10.3389/fcomp.2026.1873283
Abid, N., Qazi, O., Ahmed, H., Rafaqat, S., & Yousaf, M. (2026) · Diagnostics · doi:10.3390/diagnostics16213517
Metadata
- Authors
- Trujillo-Guerrero, M. F., Román-Niemes, S., Jaén-Vargas, M., Cadiz, A., Fonseca, R., & Serrano-Olmedo, J. J.
- Affiliations
- Universidad Politécnica de Madrid (ES)
- Universidad Yachay Tech (EC)
- Digevo, Santiago, Chile
- Center for Biomedical Technology (CTB), Universidad Politécnica de Madrid, Madrid, Spain
- ORCID
- Venue
- IEEE Access
- Published
- 2023
- Volume
- 11
- Pages
- 106650-106669
- ISSN
- 2169-3536 (ISSN-L)
- OpenAlex
- W4387010474
- Cited by
- 28
- References
- 47
- Citations by year
- 2026: 7 · 2025: 11 · 2024: 10
- FWCI
- 2.54
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- IEEE Access (publishedVersion, cc-by)
- Hispana (acceptedVersion, cc-by)
- UPM Digital Archive (Technical University of Madrid) (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Ministerio de Ciencia, Innovación y Universidades of Spain (PGC2018-097531-B-I00)
- Cloud Computing
- Microsoft Founders Hub
- 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 › Computer Science › Human-Computer Interaction › Hand Gesture Recognition Systems
- Keywords
- human activity recognition, inertial measurement unit, convolutional neural networks, LSTM, transformer models, sensor data acquisition, feature selection, window size, sampling frequency, signal correlation
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.
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.