Zhuojin Pipe Fitting Co., LtdThree-Channel Millimeter Wave Radar Target Recognition Method Study Note
Literature Overview
The paper by Wang Jiabin and colleagues, published in Modern Radar (2026, Vol. 48, No. 5, pp. 14-22), addresses a persistent challenge in automotive and surveillance radar systems: the limited recognition accuracy when using single-channel Range-Doppler (RD) spectrograms as input to Convolutional Neural Network (CNN) classifiers. The authors propose a three-channel CNN architecture that enriches feature representation through complementary signal processing stages, ultimately feeding the enhanced feature maps into a modified S-MobileNet lightweight network. This work is supported by the National Natural Science Foundation of China (Grant No. 61502228) and originates from the Ministry of Education Key Laboratory of Radar Imaging and Microwave Photonics at Nanjing University of Aeronautics and Astronautics.
Core Technical Approach
The methodology follows a structured signal-to-classification pipeline that can be decomposed into three principal stages.
Stage 1: Raw RD Spectrum Acquisition
The authors employ a two-dimensional Fast Fourier Transform (2D-FFT) combined with a Vector Mean Cancellation (VMC) algorithm to process raw radar echo signals. The 2D-FFT converts the time-frequency domain echo into the Range-Doppler domain, producing the conventional RD spectrogram. The VMC algorithm subsequently suppresses stationary clutter and reduces interference from multiple targets. This dual-processing approach ensures that the baseline RD spectrum retains meaningful target signatures while minimizing background noise.
Stage 2: Three-Channel Feature Extraction
The critical innovation lies in deriving three distinct feature channels from the processed RD spectrum:
- Channel 1 — Raw Range Information: Preserves the unprocessed distance profile of detected targets, capturing fundamental geometric separation between objects.
- Channel 2 — Interference-Reduced RD Spectrum: Obtained after applying the VMC algorithm to suppress multi-target interference, this channel emphasizes isolated target signatures.
- Channel 3 — CFAR-Processed RD Spectrum: The Constant False Alarm Rate (CFAR) detection stage highlights statistically significant targets while suppressing noise floor variations, providing a threshold-based enhancement of true target returns.
Stage 3: Lightweight CNN Classification
The three-channel feature map is input into a modified S-MobileNet architecture. The modifications focus on reducing computational complexity while preserving classification accuracy, making the network suitable for deployment in resource-constrained automotive radar platforms. The improved network achieves significant reductions in parameter count and floating-point operations per second (FLOPS) relative to baseline MobileNet variants.
Performance and Comparative Analysis
| Metric | Single-Channel Baseline | Three-Channel Proposed Method | Improvement |
|---|---|---|---|
| Recognition Accuracy | Lower in complex scenes | Significantly improved | Substantial gain |
| Feature Diversity | Single RD spectrum | Three complementary channels | Enhanced dimensionality |
| Model Complexity | Standard CNN | Modified S-MobileNet | Reduced parameters |
| Computational Load | Higher | Lower | Suitable for embedded use |
| Robustness to Clutter | Limited | Improved via CFAR and VMC | Better in multi-target scenarios |
The experimental results demonstrate that the three-input-channel approach substantially enhances target recognition accuracy and robustness in complex scenes. The combination of raw range data, interference-cancelled spectrograms, and CFAR-enhanced features provides the network with a richer representation space, enabling it to distinguish targets that would otherwise be confused in a single-channel setting.
Engineering Practice Reflections
From an engineering systems perspective, this work illustrates a broader principle applicable across technical domains: feature enrichment through multi-stage signal processing can compensate for limitations in raw data quality. In radar systems, the single-channel RD spectrogram suffers from feature homogeneity — all information is encoded in one representational space, limiting the discriminative power of downstream classifiers. By decomposing the signal into complementary processing paths and recombining them at the network input, the authors effectively increase the information entropy available to the classifier.
The choice of S-MobileNet as the backbone network is pragmatic. Automotive radar platforms operate under strict power and latency budgets, and lightweight networks are essential for real-time inference. The modifications to S-MobileNet that reduce complexity without proportionally sacrificing accuracy represent a well-calibrated engineering trade-off.
The CFAR processing stage deserves particular attention. In practice, CFAR thresholds must be calibrated to the operating environment. The authors' approach of incorporating CFAR-processed data as a separate channel rather than as a preprocessing step that replaces the original data is a sophisticated design choice — it preserves the original signal information while adding a statistical enhancement layer.
Key Questions and Study Insights
Several questions emerge from this work that warrant further investigation. First, the three-channel approach increases input dimensionality by a factor of three, which raises concerns about data storage and bandwidth in real-time systems. Second, the optimal balance between the three channels may vary with operating conditions such as target density, range, and Doppler spread. Third, the VMC algorithm's effectiveness depends on the statistical properties of the clutter environment, and its performance in highly dynamic scenes remains to be validated.
The broader implication for signal processing research is that multi-channel feature fusion, when designed with complementary processing stages, can yield disproportionate improvements in classification performance relative to the additional computational cost. This principle extends beyond radar to any domain where signal representation is constrained by single-domain processing.
This study provides a clear and well-executed demonstration that enriching input feature representation through multi-stage signal processing can substantially improve pattern recognition performance, and the proposed three-channel architecture represents a practical and deployable solution for automotive radar target classification in complex environments.