Multi-Task Self-Supervised Learning for Vehicle Classification Based on Carrier-Free UWB Radars

Abstract

Recently, target recognition based on the carrier-free ultra-wideband (UWB) radar has attracted increasing attention, as compared with narrow-band radars and other traditional UWB radars, short duration and extreme bandwidth guarantee that carrier-free UWB echoes carry richer knowledge concerning the target of interest. However, its widespread application to target recognition faces a challenge; that is, the target-aspect sensitivity issue. The target-aspect sensitivity refers to the phenomenon that carrier-free UWB echoes significantly vary as target-aspect changes, decreasing recognition accuracy. To address this problem, the paper presents a novel multi-task self-supervised learning model that can capture abundant semantic information relying on data itself instead of identity annotations. Firstly, the model is formulated as a target-aspect-invariant task, which maximizes the mutual information between original data and transformed ones to learn insensitive representations. Then given the impact of noise on recognition performance, a stacked convolutional denoising auto-encoders (SCDAE) is combined with the proposed self-supervised learning framework to extract noise-robust and target-aspect-invariant features synchronously. Extensive experiments on the measured and synthetic data demonstrate that the proposed model can achieve excellent classification performance.