Track-Oriented Self-Supervised Learning for Micro-Doppler Target Classification

Deep learning's shift from supervised to unsupervised methods has revolutionized many applications, exemplified by innovations such as DALL-E-2 and Chat-GPT. This paper focuses on extending the application of self-supervised learning, specifically in the radar domain, by introducing a simple, yet effective approach to training self-supervised models inspired by SimCLR, but tailored for micro-Doppler spectrograms. By using a target tracker we can, for a given track choose different contrastive samples by selecting spectrogram snippets from random time points along the whole track. This approach leverages the natural augmentation that occurs during a track sequence of a target. Despite its simplicity, this approach outperforms traditional augmentation techniques when training contrastive models. We demonstrated our approach by training a foundation model for radar classification and fine-tuning the model on two separate datasets. Our training approach improves classification performance significantly, and in a scenario where we have 100 times more unlabeled than labeled training samples, we boost performance with an impressive 24.79 percentage points (pp.), over a conventional supervised approach.