CAN: Classification-Based Autoencoder Network for Enhanced Out-of-Distribution Detection in Radar-Based Hand Gesture Recognition

Out-of-distribution (OOD) detection is essential for the robustness of radar-based gesture recognition systems. This paper proposes Classification-Based Autoencoder Network (CAN), as a novel OOD detection approach for radar-based hand gesture recognition. CAN uses an autoencoder to reconstruct radar gesture data, integrated with a classifier to learn a discriminative latent space, resulting in improved OOD detection performance. The evaluation of CAN is conducted on a dataset consisting of nine in-distribution gesture classes and a challenging OOD dataset containing six gestures that closely resemble the in-distribution gestures. The results indicate that CAN substantially improves the detection of challenging OOD radar samples compared to current state-of-the-art (SOTA) methods. It reduces the false positive rate (FPR) to 21%, which is significantly lower than the 58% FPR of the existing baseline. Additionally, it achieves an Area under the receiver operating characteristic curve (AUROC) of 95%, surpassing the 78% AUROC of the baseline method.