Wearable Sensor-Based Human Activity Recognition

Wearable sensor-based human activity recognition (HAR) encompasses the acquisition, processing and interpretation of physiological and kinematic signals captured by on-body devices to infer a user’s actions and context. Common sensor modalities include accelerometers, gyroscopes and magnetometers, often integrated in smartphones, smartwatches or bespoke wearable platforms. Data processing pipelines typically involve signal pre-processing, segmentation via sliding windows, feature extraction in time and frequency domains, and classification using traditional machine learning or deep learning techniques. Recent advances have seen a shift from manual feature engineering to end-to-end architectures that automatically learn hierarchical representations. Particular attention has been paid to modelling temporal dynamics through recurrent neural networks or temporal convolutional architectures. Emerging strategies such as self-supervised learning exploit vast unlabelled datasets to improve generalisability across devices and populations. Sensor fusion techniques further enhance robustness by combining multiple data streams. Applications span healthcare monitoring and rehabilitation, fall detection for older adults, fitness tracking, occupational safety and human–machine interaction. The global proliferation of low-cost wearable devices is driving scalable HAR solutions that operate in real time on edge devices, raising considerations around power consumption, privacy and model adaptability. The research community continues to address challenges of heterogeneity in sensor placement, variability in user behaviour and the scarcity of labelled data, aiming to deliver accessible and reliable activity recognition systems for diverse real-world settings.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Wearable Sensor-Based Human Activity Recognition publication trend

The graph below shows the total number of articles in wearable sensor-based human activity recognition across all publications each year (not limited to Nature Index journals).

Technical terms

Accelerometer: A device that measures linear acceleration forces along one or more axes.

Gyroscope: A sensor that detects angular velocity and orientation changes.

Self-supervised learning: A technique using automatically generated supervisory signals to learn representations from unlabelled data.

Convolutional neural network (CNN): A deep learning architecture that applies convolutional filters to extract hierarchical spatial or temporal features.

Long short-term memory (LSTM): A type of recurrent neural network cell designed to model long-term dependencies in sequence data.

Sensor fusion: The process of combining data from multiple sensors to improve the reliability and accuracy of inference.

Edge computing: Computation performed locally on a wearable or mobile device, reducing reliance on remote servers.

References

1. Self-supervised learning for human activity recognition using 700,000 person-days of wearable data. npj Digital Medicine (2024).
2. Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition. Sensors (2016).
3. Physical Human Activity Recognition Using Wearable Sensors. Sensors (2015).
4. LSTM-CNN Architecture for Human Activity Recognition. IEEE Access (2020).
5. Machine Learning Methods for Classifying Human Physical Activity from On-Body Accelerometers. Sensors (2010).
6. Window Size Impact in Human Activity Recognition. Sensors (2014).
7. Fusion of Smartphone Motion Sensors for Physical Activity Recognition. Sensors (2014).
8. Deep Recurrent Neural Networks for Human Activity Recognition. Sensors (2017).