Three-Channel Millimeter Wave Radar Target Recognition Method Study Note

Literature Overview

This paper by Wang Jiabin and colleagues from the Key Laboratory of Radar Imaging and Microwave Photonics at Nanjing University of Aeronautics and Astronautics presents a three-channel CNN-based radar target recognition method published in Modern Radar (2026, Vol. 48, No. 5). The work addresses a fundamental limitation in conventional single-channel range-Doppler (RD) spectrum processing, where feature homogeneity leads to degraded recognition accuracy in complex scenarios. The research is supported by the National Natural Science Foundation of China (Grant 61502228).

Core Technical Approach

The methodology proceeds through three sequential stages. First, the raw radar echo signal is processed using a combined two-dimensional fast Fourier transform and vector mean cancellation algorithm to obtain the original RD spectrum data. Second, three distinct feature channels are extracted from this base data: the target's original range information, the RD spectrum after multi-target interference cancellation, and the RD spectrum processed through constant false alarm rate (CFAR) detection. Third, these three-channel feature maps are fed into an improved S-MobileNet lightweight convolutional neural network for feature learning and classification.

The key innovation lies in the multi-channel feature augmentation strategy. Rather than relying on a single RD spectrum representation, the method enriches the input dimensionality by incorporating complementary information from different signal processing stages. This approach effectively increases the discriminative power of the feature space without proportionally increasing computational complexity, thanks to the lightweight S-MobileNet architecture.

Technical Points and Engineering Relevance

From a signal processing perspective, the three channels address different aspects of target characterization. The original range channel preserves raw spatial information, the interference-cancelled channel removes multi-target contamination, and the CFAR-processed channel highlights statistically significant returns. This layered feature extraction philosophy is analogous to multi-pass quality inspection in pipe manufacturing, where different non-destructive testing methods (RT, UT, MT, PT) each reveal different defect characteristics, and their combined use significantly improves overall detection reliability.

The S-MobileNet improvement maintains the lightweight structure suitable for vehicle-mounted radar applications, which is critical for real-time processing constraints in automotive and industrial sensing environments. The experimental results demonstrate that the three-input-channel method significantly improves target recognition accuracy and robustness under complex conditions while reducing model complexity and computational load.

Study Insights and Implications

The multi-channel feature fusion strategy presented here offers a transferable concept for engineering quality control systems. In steel pipe manufacturing quality assurance, single-method inspection often yields false negatives or false positives. By combining results from multiple independent inspection channels—such as ultrasonic testing for internal defects, magnetic particle testing for surface discontinuities, and dimensional inspection for geometric compliance—engineers can achieve recognition-level reliability comparable to the three-channel radar approach described in this paper. The lightweight network architecture also suggests that complex multi-input decision systems can be implemented with manageable computational resources, which has implications for edge computing in smart manufacturing environments.