Sliding residual network for high-speed target detection in additive white Gaussian noise environments

Abstract

Facing high-speed targets, the traditional radar target detection framework may suffer from performance degradation due to the signal-to-noise ratio loss caused by Doppler mismatches, especially for phase coded waveforms. In this paper, an end-to-end sliding residual network detector (SRND), which is derived from the likelihood ratio test, is proposed to detect high-speed targets in additive white Gaussian noise environments with a single radar echo pulse. The SRND uses a residual network with an efficient depth to increasingly capture the representations of target echoes, and we partially show this process through visualization. The SRND is robust to target velocities because the employed residual network utilizes layers of convolutional filters to match with target echoes of both low-speed and high-speed targets. Besides, with a waveform adapter, the SRND is compatible with different waveforms, that is to say, the SRND needs to be trained only once and then can cope with different phase modulations of waveforms. More importantly, the SRND, which is trained with computer-generated data only, can deal with not only simulated data but also measured data. Numerical experiments are given to demonstrate the superior detection performance of the SRND over the traditional detector.