TL;DR
This paper introduces E-BDL, a novel deep learning framework that tackles the challenge of "band-dependent patterns" in radar sensing data. By adaptively filtering radar spectrograms into salient sub-bands and learning representations from these sub-views, E-BDL significantly improves the accuracy of classifying human activities and physiological signals, demonstrating its potential for enhanced radar sensing applications.
What Problem Does This Paper Address?
Radar sensors offer a privacy-preserving way to monitor human movement and physiological signals by leveraging the Doppler effect. Deep learning (DL) has shown promise in analyzing this data, often by converting radar signals into time-frequency representations (TFRs), like spectrograms, which are then treated as images. However, a significant challenge arises from "band-dependent patterns": the same type of motion or physiological signal can manifest differently across various frequency bands within the TFR. This is due to factors like varying speeds of body parts or different power scales associated with different radar cross-sections. Traditional DL models, especially Convolutional Neural Networks (CNNs), can struggle with these variations, potentially overlooking weaker signals in lower-power frequency bands or being misled by translation invariance assumptions. This paper aims to develop a DL framework that can effectively learn from these band-dependent characteristics to improve radar sensing performance.
Methods
The proposed Enhanced Band-Dependent Learning (E-BDL) framework consists of three main modules:
- Adaptive Sub-band Filtering Module: This module addresses the band-dependent patterns by adaptively identifying salient sub-bands within the TFR. It uses a modified generalized Gaussian distribution to create learnable filters that scale power values at each frequency. These filters generate multiple "sub-views" of the original spectrogram, each focusing on different frequency ranges and power levels. The parameters of these filters (center frequency and bandwidth) are trainable.
- Representation Learning Module: Each sub-view generated by the filtering module is fed into a DL-based encoder. The paper uses a pretrained ResNet-18 backbone for this purpose, which is common for its strong performance in image recognition. The encoder transforms each sub-view into a low-dimensional representation vector. Crucially, all sub-views are processed by the same encoder to ensure fair comparison in the subsequent module.
- Sub-View Contrastive Module: This module aims to learn discriminative representations by leveraging relationships between the sub-views. It employs a novel "sub-view contrastive loss function" (SCLoss) that uses class proxies. This loss function encourages representations from the same class but different sub-views to be close, while pushing apart representations from different classes. This helps ensure that each sub-view captures distinct and valuable features, leading to a more robust and interpretable representation space.
Datasets
The paper evaluates E-BDL on two radar sensing datasets:
- Vital Sign Dataset: This dataset, collected from a 24 GHz continuous wave radar, captures cardiorespiratory activities from 30 healthy volunteers across five hemodynamic scenarios (resting, Valsalva, apnea, tilt-up, tilt-down). This dataset is noted for its significant class imbalance, particularly in the "resting" category.
- Gait Dataset: This dataset is derived from walking animations simulating four Alzheimer's disease (AD) and AD-related dementia (ADRD) risk levels (normal, subtle abnormality, moderate abnormality, severe abnormality). The radar used is a frequency-modulated continuous-wave radar. This dataset is described as a "large, real-world, publicly available gait dataset" for ADRD risk evaluation, though it's based on simulations.
The paper uses standard cross-validation techniques (5-fold and leave-one-out) for the vital-sign dataset and a fixed train/validation/test split for the gait dataset.
Key Insights and Ideas
- Band-Dependent Patterns are Crucial: The paper strongly emphasizes that simply treating radar spectrograms as generic images misses critical information embedded in frequency-specific patterns and power variations.
- Adaptive Filtering is Key: Instead of fixed sub-band divisions, an adaptive approach that learns the optimal frequency bands for analysis is shown to be more effective. This allows the model to focus on the most informative parts of the spectrum for a given task.
- Sub-View Contrastive Learning for Discrimination: Learning representations from multiple, filtered "sub-views" and then contrasting them against each other (and class proxies) is a powerful way to force the model to learn distinct features and improve generalization, especially when inter-class differences are subtle.
- Interpretability through Learned Filters: The learned sub-band filters themselves provide insights into which frequency ranges are most important for different tasks, offering a degree of interpretability to the DL model.
Key Results
- Vital Sign Classification: E-BDL-ResNet achieved competitive overall accuracy (0.959 in 5-fold validation) and demonstrated superior performance in subject-independent (LOO validation) settings (0.722 overall accuracy), outperforming state-of-the-art models like ConvNeXt-T and EfficientNetV2-S in generalization. It showed particular strength in capturing subtle differences in categories like resting, tilt-up, and tilt-down.
- ADRD Risk Evaluation (Gait): E-BDL-ResNet achieved the highest overall accuracy (0.943) on the gait dataset, outperforming models like BDL-ResNet (which uses uniform sub-band division) by 2%. It showed significant improvements in detecting normal and subtle abnormality gait patterns, with fewer false positives.
- Efficiency: E-BDL-ResNet uses a ResNet-18 backbone and only slightly increases parameters compared to ResNet-18, while offering improved performance.
Limitations and Future Work
- Adaptive Filter Complexity: The current adaptive sub-band filter uses a modified Gaussian distribution with two trainable parameters (location and scale). The authors suggest that the shape parameter could also be made trainable, potentially with constraints.
- Feature Fusion: The sub-views are currently handled independently in the representation learning and contrastive modules. The paper proposes that incorporating a feature fusion section to integrate knowledge from all sub-views could further enhance performance.
- Model Agnosticism: While E-BDL is designed to be model-agnostic, its performance was validated using ResNet-18. Future work could explore using other advanced encoders.
- Real-time Deployment: The authors believe E-BDL has significant potential for implementation on edge radar computing equipment for real-time detection.
Summary updated