A Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition

Chang, Youngjae; Mathur, Akhil; Isopoussu, Anton; Song, Junehwa; Kawsar, Fahim

doi:10.1145/3380985

Cited by 90 publications

(63 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Qin et al [32] propose Adaptive Spatial-Temporal Transfer Learning (ASTTL) to allow more accurate source selection to perform domain adaption. Chang et al [4] have looked into feature matching and adversarial learning in adapting the activity model from one sensor position to another. These recent techniques are built on a similar assumption that both source and target domains share the same feature space so that they can share the same activity model [6] or feature extractor [4], [32].…”

Section: A Domain Adaptation In Human Activity Recognitionmentioning

confidence: 99%

“…Chang et al [4] have looked into feature matching and adversarial learning in adapting the activity model from one sensor position to another. These recent techniques are built on a similar assumption that both source and target domains share the same feature space so that they can share the same activity model [6] or feature extractor [4], [32]. Domain adaptation on binary sensor data is different from accelerometer data as the main challenge is to tackle heterogeneity of feature spaces.…”

Section: A Domain Adaptation In Human Activity Recognitionmentioning

confidence: 99%

“…Domain adaptation techniques have been increasingly applied to HAR applications; for example on accelerometer data, much effort has been devoted to transferring activity knowledge learned on one sensor position (e.g., arm) to another position (e.g., leg) [4], [6], [20]. These techniques often work well as they assume that the source and target domains share feature space and they only need to tackle the difference in distributions.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

Sanabria

Zambonelli

2021

IEEE Access

View full text Add to dashboard Cite

Sensor-based human activity recognition (HAR) is having a significant impact in a wide range of applications in smart city, smart home, and personal healthcare. Such wide deployment of HAR systems often faces the annotation-scarcity challenge; that is, most of the HAR techniques, especially the deep learning techniques, require a large number of training data while annotating sensor data is very time-and effort-consuming. Unsupervised domain adaptation has been successfully applied to tackle this challenge, where the activity knowledge from a well-annotated domain can be transferred to a new, unlabelled domain. However, these existing techniques do not perform well on highly heterogeneous domains. This article proposes shift-GAN that integrate bidirectional generative adversarial networks (Bi-GAN) and kernel mean matching (KMM) in an innovative way to learn intrinsic, robust feature transfer between two heterogeneous domains. Bi-GAN consists of two GANs that are bound by a cyclic constraint, which enables more effective feature transfer than a classic, single GAN model. KMM is a powerful non-parametric technique to correct covariate shift, which further improves feature space alignment. Through a series of comprehensive, empirical evaluations, shift-GAN has not only achieved its superior performance over 10 state-of-the-art domain adaptation techniques but also demonstrated its effectiveness in learning activity-independent, intrinsic feature mappings between two domains, robustness to sensor noise, and less sensitivity to training data.INDEX TERMS Human activity recognition, domain adaptation, ensemble learning, generative adversarial networks, covariate shift, kernel mean matching.

show abstract

Section: A Domain Adaptation In Human Activity Recognitionmentioning

confidence: 99%

Section: A Domain Adaptation In Human Activity Recognitionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

Sanabria

Zambonelli

2021

IEEE Access

View full text Add to dashboard Cite

show abstract

“…However, such techniques are hard to apply to our problem domain since this still requires enough real training data (atleast in unlabelled form) from IMU to achieve sufficient convergence of the domain adaptation process. Furthermore, each user's finger motion pattern as well as natural variations in sensor wearing positions could lead to different distributions in the sensor data [15,20] thus entailing more training data for each setting. On the other hand, ZeroNet performs comparable to models developed with semi-supervised domain adaptation [27,83] which need partial labelled real IMU data and even outperforms models fully trained on our own real IMU dataset (details in Sec.…”

Section: Introductionmentioning

confidence: 99%

When Video meets Inertial Sensors

Liu

Zhang

Gowda

2021

Proceedings of the International Conference on Internet-of-Things Design and Implementation

View full text Add to dashboard Cite

Ubiquitous finger motion tracking enables a number of exciting applications in augmented reality, sports analytics, rehabilitationhealthcare etc. While finger motion tracking with cameras is very mature, largely due to availability of massive training datasets, there is a dearth of training data for developing robust machine learning (ML) models for wearable IoT devices with Inertial Measurement Unit (IMU) sensors. Towards addressing this problem, this paper presents ZeroNet, a system that shows the feasibility of developing ML models for IMU sensors with zero training overhead. ZeroNet harvests training data from publicly available videos for performing inferences on IMU. The difference in data among video and IMU domains introduces a number of challenges due to differences in sensor-camera coordinate systems, body sizes of users, speed/orientation changes during gesturing, sensor position variations etc. ZeroNet addresses these challenges by systematically extracting motion data from videos and transforming them into acceleration and orientation information measured by IMU sensors. Furthermore, data-augmentation techniques are exploited that create synthetic variations in the harvested training data to enhance the generalizability and robustness of the ML models to user diversity. Evaluation with 10 users demonstrates a top-1 accuracy of 82.4% and a top-3 accuracy of 94.8% for recognition of 50 finger gestures thus indicating promise. While we have only scratched the surface, we outline a number of interesting possibilities for extending this work in the cross-disciplinary areas of computer vision, machine learning, and wearable IoT for enabling novel applications in finger motion tracking. CCS CONCEPTS• Human-centered computing → Mobile devices; Ubiquitous and mobile computing design and evaluation methods; • Computing methodologies → Neural networks.

show abstract

“…In a recent previous work [8], the authors developed and evaluated three adaptation techniques for HAR: data augmentation, feature matching and confusion maximization. They concluded that performing unsupervised domain adaptation and domain generalization for HAR in wearing diversity problem is a very complicated task, and they proposed some practical considerations for unsupervised domain adaptation.…”

Section: Introductionmentioning

confidence: 99%

Adaptation and Selection Techniques Based on Deep Learning for Human Activity Recognition Using Inertial Sensors

Gil-Martín¹,

Antúnez-Durango²,

San-Segundo³

2020

7th International Electronic Conference on Sensors and Applications

View full text Add to dashboard Cite

Deep learning techniques have been widely applied to Human Activity Recognition (HAR), but a specific challenge appears when HAR systems are trained and tested with different subjects. This paper describes and evaluates several techniques based on deep learning algorithms for adapting and selecting the training data used to generate a HAR system using accelerometer recordings. This paper proposes two alternatives: autoencoders and Generative Adversarial Networks (GANs). Both alternatives are based on deep neural networks including convolutional layers for feature extraction and fully-connected layers for classification. Fast Fourier Transform (FFT) is used as a transformation of acceleration data to provide an appropriate input data to the deep neural network. This study has used acceleration recordings from hand, chest and ankle sensors included in the Physical Activity Monitoring Data Set (PAMAP2) dataset. This is a public dataset including recordings from nine subjects while performing 12 activities such as walking, running, sitting, ascending stairs, or ironing. The evaluation has been performed using a Leave-One-Subject-Out (LOSO) cross-validation: all recordings from a subject are used as testing subset and recordings from the rest of the subjects are used as training subset. The obtained results suggest that strategies using autoencoders to adapt training data to testing data improve some users’ performance. Moreover, training data selection algorithms with autoencoders provide significant improvements. The GAN approach, using the generator or discriminator module, also provides improvement in selection experiments.

show abstract

A Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition

Cited by 90 publications

References 27 publications

Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

Unsupervised Domain Adaptation in Activity Recognition: A GAN-Based Approach

When Video meets Inertial Sensors

Adaptation and Selection Techniques Based on Deep Learning for Human Activity Recognition Using Inertial Sensors

Contact Info

Product

Resources

About