Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On Input Formats for Radar Micro-Doppler Signature Processing by Convolutional Neural Networks

Mikolaj Czerkawski, Carmine Clemente, Craig Michie, Christos Tachtatzis

TL;DR

The paper tackles how the Doppler-time input format, including phase information, impacts CNN-based classification of radar Micro-Doppler signatures. It systematically compares multiple formats (magnitude, phase with wrap and unwrap, real/imag, and hybrids) using both single-domain and a novel multi-domain CNN architecture with representation-specific encoders. It demonstrates that unwrapped phase significantly boosts performance and that multi-domain fusion with a simple meta-classifier can push accuracy to about $0.947$, highlighting the value of leveraging phase information. The findings offer practical guidance for designing input representations in radar ATR systems and establish a framework for cross-format evaluation across datasets.

Abstract

Convolutional neural networks have often been proposed for processing radar Micro-Doppler signatures, most commonly with the goal of classifying the signals. The majority of works tend to disregard phase information from the complex time-frequency representation. Here, the utility of the phase information, as well as the optimal format of the Doppler-time input for a convolutional neural network, is analysed. It is found that the performance achieved by convolutional neural network classifiers is heavily influenced by the type of input representation, even across formats with equivalent information. Furthermore, it is demonstrated that the phase component of the Doppler-time representation contains rich information useful for classification and that unwrapping the phase in the temporal dimension can improve the results compared to a magnitude-only solution, improving accuracy from 0.920 to 0.938 on the tested human activity dataset. Further improvement of 0.947 is achieved by training a linear classifier on embeddings from multiple-formats.

On Input Formats for Radar Micro-Doppler Signature Processing by Convolutional Neural Networks

TL;DR

The paper tackles how the Doppler-time input format, including phase information, impacts CNN-based classification of radar Micro-Doppler signatures. It systematically compares multiple formats (magnitude, phase with wrap and unwrap, real/imag, and hybrids) using both single-domain and a novel multi-domain CNN architecture with representation-specific encoders. It demonstrates that unwrapped phase significantly boosts performance and that multi-domain fusion with a simple meta-classifier can push accuracy to about , highlighting the value of leveraging phase information. The findings offer practical guidance for designing input representations in radar ATR systems and establish a framework for cross-format evaluation across datasets.

Abstract

Convolutional neural networks have often been proposed for processing radar Micro-Doppler signatures, most commonly with the goal of classifying the signals. The majority of works tend to disregard phase information from the complex time-frequency representation. Here, the utility of the phase information, as well as the optimal format of the Doppler-time input for a convolutional neural network, is analysed. It is found that the performance achieved by convolutional neural network classifiers is heavily influenced by the type of input representation, even across formats with equivalent information. Furthermore, it is demonstrated that the phase component of the Doppler-time representation contains rich information useful for classification and that unwrapping the phase in the temporal dimension can improve the results compared to a magnitude-only solution, improving accuracy from 0.920 to 0.938 on the tested human activity dataset. Further improvement of 0.947 is achieved by training a linear classifier on embeddings from multiple-formats.
Paper Structure (12 sections, 4 figures, 5 tables)

This paper contains 12 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Potential Formats
  5. Processing Complex Data by Neural Networks
  6. Testing Methods
  7. Results
  8. Training Details
  9. Classification Test (Single-Domain)
  10. Classification Test (Multi-Domain)
  11. Further Analysis
  12. Conclusion

Figures (4)

  • Figure 1: Single-channel formats considered in analysis.
  • Figure 2: Diagram of the employed convolutional neural network architecture. The variable in_ch indicates the number of input channels that changes depending on the input representation format (either 1, 2, or 4 channels, as described in the text).
  • Figure 3: Multi-domain classifier training, where only 2 domains are used at any training step.
  • Figure 4: Illustration of saliency map for each input representation and each class. A threshold of 25% of the maximum value is applied for visualization.