Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Novel Micro-Doppler Coherence Loss for Deep Learning Radar Applications

Mikolaj Czerkawski, Christos Ilioudis, Carmine Clemente, Craig Michie, Ivan Andonovic, Christos Tachtatzis

TL;DR

This work addresses the mismatch between conventional loss functions and the unique features of micro-Doppler radar signals by introducing a micro-Doppler coherence loss that promotes spectral alignment within Doppler bands. The loss is formulated by transforming time-domain representations into Doppler-cadence maps using $\mathcal{F}_t$, normalizing them, and minimizing $\mathcal{L}_{\mu\mathrm{D}}(y,\hat{y})$ alongside the standard $\mathcal{L}_{\mathrm{MSE}}$, yielding $J_\theta(y,\hat{y}) = \mathcal{L}_{\mathrm{MSE}}(y,\hat{y}) + \beta\cdot \mathcal{L}_{\mu\mathrm{D}}(y,\hat{y})$. The method is validated on real human activity radar data, showing that the augmented objective improves robustness to additive noise in both unsupervised pre-training and downstream classification, with notable gains such as accuracy increasing from $0.48$ to $0.61$ at $-5$ dB SNR and an additional $\sim1.2$ dB tolerance. The results suggest that prioritizing micro-Doppler features via coherence loss yields more resilient models for micro-Doppler analysis and can generalize to various radar learning tasks.

Abstract

Deep learning techniques are subject to increasing adoption for a wide range of micro-Doppler applications, where predictions need to be made based on time-frequency signal representations. Most, if not all, of the reported applications focus on translating an existing deep learning framework to this new domain with no adjustment made to the objective function. This practice results in a missed opportunity to encourage the model to prioritize features that are particularly relevant for micro-Doppler applications. Thus the paper introduces a micro-Doppler coherence loss, minimized when the normalized power of micro-Doppler oscillatory components between input and output is matched. The experiments conducted on real data show that the application of the introduced loss results in models more resilient to noise.

A Novel Micro-Doppler Coherence Loss for Deep Learning Radar Applications

TL;DR

This work addresses the mismatch between conventional loss functions and the unique features of micro-Doppler radar signals by introducing a micro-Doppler coherence loss that promotes spectral alignment within Doppler bands. The loss is formulated by transforming time-domain representations into Doppler-cadence maps using , normalizing them, and minimizing alongside the standard , yielding . The method is validated on real human activity radar data, showing that the augmented objective improves robustness to additive noise in both unsupervised pre-training and downstream classification, with notable gains such as accuracy increasing from to at dB SNR and an additional dB tolerance. The results suggest that prioritizing micro-Doppler features via coherence loss yields more resilient models for micro-Doppler analysis and can generalize to various radar learning tasks.

Abstract

Deep learning techniques are subject to increasing adoption for a wide range of micro-Doppler applications, where predictions need to be made based on time-frequency signal representations. Most, if not all, of the reported applications focus on translating an existing deep learning framework to this new domain with no adjustment made to the objective function. This practice results in a missed opportunity to encourage the model to prioritize features that are particularly relevant for micro-Doppler applications. Thus the paper introduces a micro-Doppler coherence loss, minimized when the normalized power of micro-Doppler oscillatory components between input and output is matched. The experiments conducted on real data show that the application of the introduced loss results in models more resilient to noise.
Paper Structure (7 sections, 4 equations, 4 figures)

This paper contains 7 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Proposed Solution
  4. Evaluation
  5. Unsupervised Stage
  6. Classification
  7. Conclusions

Figures (4)

  • Figure 1: Diagram of the used hybrid model.
  • Figure 2: Comparison of autoencoding loss curves. (a) Reconstruction Loss (b) micro-Doppler Loss and (c) scatter plot of the Reconstruction Loss Change and micro-Doppler Loss Change
  • Figure 3: Comparison of the relationship between accuracy and of tested samples
  • Figure 4: Comparison of obtained confusion matrix at -5 $\mathrm{SNR}_{\mathrm{dB}}$ for (a) a scenario with no micro-Doppler loss applied (b) a scenario with micro-Doppler loss backpropagation.