An Adaptive S-Method to Analyze Micro-Doppler Signals for Human Activity Classification

Fangmin Li, Chao Yang, Yuqing Xia, Xiaolin Ma, Tao Zhang, Zhou Zhou2017 · DOI

TL;DR

This paper introduces an adaptive S-method (AS-method) for analyzing micro-Doppler signals from radar, improving time-frequency resolution and suppressing interference. This enhanced analysis allows for more accurate classification of human activities, achieving 95.4% accuracy in distinguishing six different motion states.

What Problem Does This Paper Address?

Accurately identifying and classifying human activities is crucial for security and surveillance systems. Radar technology offers advantages like all-weather operation and the ability to penetrate clothing, making it a promising tool for non-contact human detection. However, existing methods for analyzing the subtle micro-Doppler signals generated by human movement often struggle with interference, noise, and achieving sufficient time-frequency resolution, especially when multiple motion components are present. This makes precise classification of activities like walking, running, or even subtle movements like arm swings challenging.

Methods

The core contribution is the Adaptive S-method (AS-method), an extension of the multiwindow S-method. The AS-method leverages orthogonal Hermite functions for better time-frequency resolution. Crucially, it adaptively adjusts window lengths based on signal characteristics to suppress cross-terms (interference between different signal components) and enhance auto-terms (the actual signal of interest). This adaptive approach aims to strike a better balance between signal concentration and interference suppression.

Following time-frequency analysis, the paper employs a dynamic threshold cutting method to extract the pure micro-Doppler signal by filtering out background noise. Feature extraction involves analyzing various characteristics of the trunk and arm movements, including Doppler frequency, bandwidth, root mean square error (RMSE) of envelopes, and arm swing period. Finally, a Support Vector Machine (SVM) classifier is trained on these extracted features to perform the activity classification.

Datasets

The paper utilizes a dataset collected by the authors using a continuous wave radar system. They describe collecting data from 12 human subjects performing seven different activities. For training and validation, they collected 600 samples, with 100 samples per motion state, from five individuals (two males, three females) performing six groups of movements. The experiments were conducted in both indoor (enclosed rooms) and outdoor (corridors) environments. No specific public benchmark dataset is mentioned as being used or introduced.

Key Insights and Ideas

  • Adaptive Windowing is Key: The AS-method's adaptive windowing, adjusting based on signal properties, is a significant improvement over fixed-window techniques for complex micro-Doppler signals. This allows for better separation of overlapping motion components.
  • Hermite Functions for Resolution: The use of Hermite functions provides a strong theoretical basis for achieving good time-frequency resolution, which is essential for distinguishing subtle motion differences.
  • Dynamic Thresholding for Clean Signals: A dynamic thresholding approach, rather than a fixed one, is more robust to varying environmental noise conditions, leading to cleaner signal extraction.
  • Comprehensive Feature Set: The proposed 12 characteristic parameters, covering direct, statistical, and relative motion characteristics, provide a rich representation of human activity, simplifying the classification task.
  • SVM's Effectiveness: The SVM classifier, when trained on these features, demonstrates high accuracy, suggesting that the chosen features effectively capture the discriminative information for activity classification.

Key Results

The AS-method processing significantly improved time-frequency aggregation, as evidenced by lower Renyi entropy and higher M values compared to standard spectrograms and Hermite multiwindow processing. The paper reports a classification accuracy of 95.4% for distinguishing six motion states (sitting still, crawling, walking with both arms, walking with a single arm, walking without arms, and running) using the trained SVM. The confusion matrix shows high accuracy for most states, with crawling achieving 100% and sitting still and running close behind. Walking states, particularly walking with a single arm, showed slightly lower accuracy due to similarities with other activities. The study also demonstrated comparable recognition accuracy in both indoor (94.1%) and outdoor (92.25%) environments.

Limitations and Future Work

The authors acknowledge that distinguishing between similar walking states (e.g., walking with both arms vs. single arm) remains a challenge, leading to lower accuracy for these categories. They also note that some characteristics (e.g., 4, 8, and 11 in Table 3) are difficult to distinguish solely by threshold cutting, necessitating the use of more advanced classification algorithms like SVM.

Future work could explore more advanced feature extraction techniques, potentially incorporating deep learning approaches, to further improve the discrimination of subtle motion differences. Investigating the performance of the AS-method with different radar systems or in more complex, cluttered environments would also be valuable. Additionally, expanding the range of activities classified and exploring real-time implementation would be important next steps.

Summary updated

BibTeX
@misc{li2017adaptive,
  author = {Fangmin Li and Chao Yang and Yuqing Xia and Xiaolin Ma and Tao Zhang and Zhou Zhou},
  title = {{An Adaptive S-Method to Analyze Micro-Doppler Signals for Human Activity Classification}},
  year = {2017},
  doi = {10.3390/s17122769},
}
Figure