
Semi-Supervised Attention-Augmented Autoencoder for Radar-Based Human Activity Recognition
Learning more from less, how we taught radar to recognize human movement using fewer labeled examples.
Summary
This work, presented at EUSIPCO 2023, proposes a semi-supervised convolutional autoencoder (CAE) enhanced with an attention mechanism for radar-based human activity recognition (HAR).
The model learns from both labeled and unlabeled radar data, making it highly data-efficient, ideal for healthcare monitoring systems where labeling is costly or impractical.
Our proposed Attention-Augmented Convolutional Autoencoder (AA-CAE) jointly trains its encoder, attention, and classifier components using a new optimization strategy. This allows global and local motion patterns to be learned simultaneously, leading to superior recognition accuracy compared to the traditional two-step semi-supervised methods.
Motivation
Radar sensors can capture fine movement patterns through micro-Doppler signatures, offering a privacy-preserving alternative to cameras.
However, labeling radar data for every activity (like walking, sitting, or falling) is tedious and time-consuming.
We aimed to reduce dependency on labeled samples by designing a joint semi-supervised framework that still maintains high accuracy, even when a large portion of the training data is unlabeled.
Model Architecture
The Attention-Augmented CAE (AA-CAE) combines three main components:
- Encoder-Decoder (CAE): Learns compact latent representations of radar micro-Doppler images.
- Attention Mechanism: Captures global context to complement the encoder’s localized features.
- Classifier: Maps the combined attention and convolutional features to human activity classes.
In contrast to the conventional two-step training, where the encoder is first pre-trained and then frozen, our joint training approach updates all components together using a combined reconstruction-plus-classification loss.

Encoder, attention, and classifier are trained simultaneously using both labeled and unlabeled data.
Semi-Supervised Joint Training
We introduce a Split-Bregman optimization approach to enable backpropagation through both the supervised and unsupervised loss components.
The joint loss function combines:
- Reconstruction loss (for all samples), and
- Prediction loss (for labeled samples).
This design ensures the encoder learns representations that directly benefit the classification task, effectively bridging representation learning and supervised learning.
Dataset
We used the University of Glasgow Human Activity Radar Dataset, captured using a 5.8 GHz FMCW radar.
Six activities were studied:
- Walking
- Sitting down
- Standing up
- Drinking water
- Bending to pick up an object
- Falling

Each radar signature encodes subtle frequency shifts corresponding to human motion.
A total of 570 micro-Doppler spectrograms (95 per activity) from 33 subjects were analyzed after cropping and normalization (76 × 76 pixels, grayscale 0-255).
Results
We compared our jointly trained semi-supervised AA-CAE with the traditional two-step training method across different training/testing splits and proportions of labeled data.
Training/Test Split | Training Method | Accuracy (%) | Std Dev |
---|---|---|---|
20/80 | Two-Step (Disjoint) | 71.13 | 0.016 |
20/80 | Joint (Proposed) | 76.43 | 0.016 |
50/50 | Two-Step (Disjoint) | 77.43 | 0.017 |
50/50 | Joint (Proposed) | 81.78 | 0.015 |
80/20 | Two-Step (Disjoint) | 80.39 | 0.016 |
80/20 | Joint (Proposed) | 85.78 | 0.016 |
The joint training consistently improved both accuracy and stability.
It achieved the same performance as the fully supervised model with roughly 20% fewer labeled samples.
Key Insights
- Joint semi-supervised learning effectively unifies representation and classification objectives.
- The attention mechanism enables the network to focus on the most informative regions in radar spectrograms.
- Achieves robust recognition even with limited labels, vital for clinical and ambient monitoring use cases.
- Improves model stability and generalization compared to two-step training.
Applications
This work supports radar-based remote patient monitoring, fall detection, and mobility tracking, areas where data privacy, safety, and continuous monitoring are critical.
By leveraging unlabeled data, the proposed framework brings us closer to scalable, real-world deployment of intelligent radar healthcare systems.
Citation
Sadeghi-Adl, Z.; Ahmad, F.
“Semi-Supervised Convolutional Autoencoder With Attention Mechanism for Activity Recognition.”
Proceedings of EUSIPCO 2023.
Temple University, Philadelphia, PA.