Recognition for Radar with Low Signal-to-Noise Ratio

Abstract

To overcome the limited identification precision of wheeled vehicles from linear frequency-modulated continuous wave radar echoes amid weak signal-to-noise ratio scenarios, this study develops a target recognition approach that integrates ICEEMDAN processing and upgraded EfficientNetV2-S architecture. First, an improved sawtooth modulation scheme for LFMCW radar is developed by introducing a constant-frequency band, and a mathematical micro-Doppler model of wheeled vehicles under this radar detection mode is established. Second, ICEEMDAN is utilized to decompose the radar echo signals, and noise reduction is realized with an improved wavelet threshold algorithm. After denoising, we perform Short-Time Fourier Transform (STFT) on the processed radar signals to generate time-frequency representations. A dedicated micro-Doppler enhancement module is then introduced to refine the visual quality of the resulting time-frequency figures. Finally, a modified EfficientNetV2-S model is constructed to achieve precise recognition of wheeled vehicle targets for radar systems. Comparative experimental tests validate that the devised algorithm yields better identification precision of vehicle objects and exhibits improved stability when facing low-SNR measurement scenarios.