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

The DOI matches its Crossref record; 234 of the 235 references it cites were found in a registry or an index; 107 of the 168 sources its claims cite were checked for support.

Past, Present, and Future of Sensor-based Human Activity Recognition Using Wearables: A Surveying Tutorial on a Still Challenging Task

Open access
Haresamudram, Harish, Tang, Chi Ian, Suh, Sungho, Lukowicz, Paul, Plötz, Thomas · 2025 · Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies

Abstract

A comprehensive survey and tutorial on sensor-based human activity recognition from heuristic handcrafted feature engineering through deep learning, self-supervised learning, and multimodal foundation models, outlining benchmark datasets, placement challenges, and open problems.

Overview

CoChat Summary

TL;DR

This surveying tutorial reviews sensor-based wearable human activity recognition from handcrafted features through deep learning and self-supervised representation learning to emerging foundational-model approaches. It argues that HAR remains challenging due to signal ambivalence, input variability, information coding mismatch, and scarce labeled data, and provides a hands-on tutorial with code and Motionsense experiments.

Key findings

  • The paper identifies four broad HAR challenges: signal ambivalence, variability of the input space, information coding, and lack of labeled training data.
  • In the tutorial experiments on Motionsense, the ECDF-RF baseline achieved a test F1-score of 81.84%, the supervised Conv.classifier achieved 85.1%, and SimCLR with frozen pre-trained encoder weights achieved 83.17%.
  • Class imbalance is substantial in wearable datasets: the null class comprises approx. 75% of the Opportunity benchmark, >35% of Mobiact is “Standing,” and transitionary activities such as “Sit to stand” comprise <5%.
  • The paper concludes there is no “one-size-fits-all” HAR method, but self-supervised representation learning is currently the predominant approach and foundational models may drive a third paradigm shift.

Methods

  • Design: a surveying tutorial that reviews the field from the Activity Recognition Chain and handcrafted features to end-to-end deep learning, self-supervised learning, data generation/augmentation, and foundational models.
  • Data and tutorial experiments: the hands-on tutorial uses the Motionsense dataset, accelerometer data only, with six locomotion-style activities from 24 participants; splits are 15 train, 4 validation, and 5 test users.
  • Main analysis: the tutorial compares distribution-based ECDF features with a Random Forest classifier, a supervised convolutional classifier, and SimCLR self-supervised learning, using mean/unweighted F1-score because it is more resistant to class imbalance.

Limitations

  • The paper states that there is no universally applicable HAR method, that large-scale richly annotated datasets are lacking, and that plug-and-play sensor-language models remain unsolved; it also notes that many sensor-language works train and evaluate on different splits of the same dataset, leaving cross-distribution performance unclear.
  • The paper states limitations for specific methods: GAN-based generation is dataset- and label-specific, requires substantial labeled data, and can suffer mode collapse; adversarial learning lacks a mechanism to measure generalization quality during training; diffusion models require considerable computational steps; and early simulation approaches often rely on MoCap and exclude gyroscope data.
  • Evident from the text: the tutorial uses only accelerometer data to reduce training time and evaluates on Motionsense’s relatively simple locomotion activities; the ECDF-RF result shows 100% train performance versus 81.84% test F1, which the paper itself notes indicates potential overfitting.

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

Integrity signals

Registry

Crossref · 10.1145/3729467

Cites

234 of 235 references verified0 citation errors · 1 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Sit push-ups and triceps dips as examples of activities that can be easily distinguished visually (as illustrated by the GPT-4o query on the included images) while producing very similar acceleration signals requiring the interpretation of subtle movements for recognition. This also illustrates the variability of the signal across different modalities.
Fig. 1 Sit push-ups and triceps dips as examples of activities that can be easily distinguished visually (as illustrated by the GPT-4o query on the included images) while producing very similar acceleration signals requiring the interpretation of subtle movements for recognition. This also illustrates the variability of the signal across different modalities.CC-BY · From the PDF
Fig. 2 An overview of the self-supervised learning pipeline. Reproduced with permission from [81].
Fig. 2 An overview of the self-supervised learning pipeline. Reproduced with permission from [81].CC-BY · From the PDF
Image 3
CC-BY · From the PDF
Fig. 3 Contrastive learning adaptations in human activity recognition. A comparison is drawn among SimCLR [184], CPC [80], BYOL [81] and SimSiam [81]. A high degree of commonality can be found among these frameworks, especially in the use of augmented anchors as positive samples for SimCLR, BYOL, and SimSiam. The CPC differs from the rest by using future samples instead of augmented views. SimCLR and CPC leverage time-misaligned samples as negatives, while BYOL and SimSiam leverage additional mechanisms to remove the requirement of negative samples.
Fig. 3 Contrastive learning adaptations in human activity recognition. A comparison is drawn among SimCLR [184], CPC [80], BYOL [81] and SimSiam [81]. A high degree of commonality can be found among these frameworks, especially in the use of augmented anchors as positive samples for SimCLR, BYOL, and SimSiam. The CPC differs from the rest by using future samples instead of augmented views. SimCLR and CPC leverage time-misaligned samples as negatives, while BYOL and SimSiam leverage additional mechanisms to remove the requirement of negative samples.CC-BY · From the PDF
Fig. 4 An overview of multi-device and multi-modal contrastive learning frameworks in human activity recognition. A comparison is drawn among ColloSSL [93], Learning from the Best (LftB) [55], COCOA [40] and SimCLR [184] (as a singledevice single-modality reference). Instead of relying on augmentations (as in SimCLR), these multi-device and multi-modal approaches leverage data from time-aligned data from different devices (ColloSSL and LftB), and different modalities (COCOA) for positive samples. Time-misaligned samples as negatives is a common feature among these approaches, with different frameworks imposing additional limitations on the sampling algorithm.
Fig. 4 An overview of multi-device and multi-modal contrastive learning frameworks in human activity recognition. A comparison is drawn among ColloSSL [93], Learning from the Best (LftB) [55], COCOA [40] and SimCLR [184] (as a singledevice single-modality reference). Instead of relying on augmentations (as in SimCLR), these multi-device and multi-modal approaches leverage data from time-aligned data from different devices (ColloSSL and LftB), and different modalities (COCOA) for positive samples. Time-misaligned samples as negatives is a common feature among these approaches, with different frameworks imposing additional limitations on the sampling algorithm.CC-BY · From the PDF
Image 6
CC-BY · From the PDF
Fig. 7 An overview of contrastive learning (CL). Data samples (images or sensor signals depending on the target task) are first passed through an encoder to obtain embeddings in a latent space. These embeddings are pushed closer to each other or pulled apart depending on whether the embedding forms a positive pair or a negative pair with the anchor.
Fig. 7 An overview of contrastive learning (CL). Data samples (images or sensor signals depending on the target task) are first passed through an encoder to obtain embeddings in a latent space. These embeddings are pushed closer to each other or pulled apart depending on whether the embedding forms a positive pair or a negative pair with the anchor.CC-BY · From the PDF
Fig. 8 Class composition of the Motionsense dataset.
Fig. 8 Class composition of the Motionsense dataset.CC-BY · From the PDF
Fig. 9 An overview of the SimCLR training pipeline for human activity recognition [184].
Fig. 9 An overview of the SimCLR training pipeline for human activity recognition [184].CC-BY · From the PDF
Image 10
CC-BY · From the PDF
Fig. 11 HAR performance obtained by the three types of representations
Fig. 11 HAR performance obtained by the three types of representationsCC-BY · From the PDF
Fig. 12 Confusion matrices for performing HAR using the three types of representations
Fig. 12 Confusion matrices for performing HAR using the three types of representationsCC-BY · From the PDF

Verification

Claims the paper makes

62 Valid · 7 Not valid · 36 Not checked · 98 Own results and interpretations

101 more claims · 26 with their cited sources checked

Create a free account to see all 203 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 47
  • Probably not a substantial claim. 59
  • The quote only partly states this claim. 1
  • Borderline: one of the tests for a substantial claim is unclear. 1

Create a free account to see these claims.

Claims other papers make about this one

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

Self-check

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

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

3 of 3

Citations

Cites

Show all 156 references

Research works printed but not linked: 14

  1. 10.5555/1622407.1622416 · doi:10.5555/1622407.1622416
  2. Spyros Gidaris, Andrei Bursuc, Jean-Baptiste Alayrac, Adrià Recasens, Mathilde Caron, Olivier Hénaff, Aäron van den Oord, and Relja Arandjelović. 2021. CVPR 2021 Tutorial on Self-Supervised Representation Learning --- gidariss.github.io. https://gidariss.github.io/self-supervised-learning-cvpr2021/. [Accessed 28-10-2024].
  3. Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016. Deep Learning. MIT Press. http://www.deeplearningbook.org.
  4. Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 6840--6851. https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf
  5. Yifan Hu. 2023. BSDGAN: Balancing Sensor Data Generative Adversarial Networks for Human Activity Recognition. In 2023 International Joint Conference on Neural Networks (IJCNN). IEEE, 1--8.
  6. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012).
  7. Lin Liao, Dieter Fox, and Henry Kautz. 2005. Location-based activity recognition. Advances in neural information processing systems 18 (2005).
  8. Hiroki Ohashi, M Al-Nasser, Sheraz Ahmed, Takayuki Akiyama, Takuto Sato, Phong Nguyen, Katsuyuki Nakamura, and Andreas Dengel. 2017. Augmenting wearable sensor data with physical constraint for DNN-based human-action recognition. In ICML 2017 times series workshop. 6--11.
  9. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 8748--8763.
  10. Milyun Ni'ma Shoumi and Sozo Inoue. 2024. Leveraging the Large Language Model for Activity Recognition: A Comprehensive Review. International Journal of Activity and Behavior Computing 2024, 2 (2024), 1--27.
  11. Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. In International Conference on Learning Representations. https://openreview.net/forum?id=St1giarCHLP
  12. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017).
  13. Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. 2022. GeoDiff: A Geometric Diffusion Model for Molecular Conformation Generation. In International Conference on Learning Representations. https://openreview.net/forum?id=PzcvxEMzvQC
  14. Jinsung Yoon, Daniel Jarrett, and Mihaela Van der Schaar. 2019. Time-series generative adversarial networks. Advances in neural information processing systems 32 (2019).

Cited by

See all 33

Metadata

Authors
Haresamudram, H., Tang, C. I., Suh, S., Lukowicz, P., & Plötz, T.
Affiliations
  • Georgia Institute of Technology (US)
  • Nokia (United Kingdom) (GB)
  • Korea University (KR)
  • German Research Centre for Artificial Intelligence (DE)
Venue
Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies
Published
2025-06-09
Volume
9
Issue
2
Pages
1-44
ISSN
2474-9567 (ISSN-L)
OpenAlex
W4411403377
Cited by
33
References
156
Citations by year
2026: 26 · 2025: 7
FWCI
11.47
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • Horizon 2020 Framework Programme (52026 HumanE- AI-Net project)
  • NSF (IS2112633)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Health Sciences › Medicine › Public Health, Environmental and Occupational Health › Physical Activity and Health
  • Physical Sciences › Engineering › Biomedical Engineering › Gait Recognition and Analysis
Keywords
human activity recognition, wearable sensors, sensor-based activity recognition, wearable computing, feature engineering, foundation models, benchmark performance

About these checks

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

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

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

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

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

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

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

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

How to cite

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

This page
CoChat. (2026). Citation check: Past, Present, and Future of Sensor-based Human Activity Recognition Using Wearables: A Surveying Tutorial on a Still Challenging Task. https://app.cochat.ai/papers/10.1145/3729467
The paper
Haresamudram, H., Tang, C. I., Suh, S., Lukowicz, P., & Plötz, T. (2025). Past, Present, and Future of Sensor-based Human Activity Recognition Using Wearables: A Surveying Tutorial on a Still Challenging Task. Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies. https://doi.org/10.1145/3729467

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.