Radar has emerged as a promising sensing modality for sign language recognition in privacy-sensitive scenarios, yet frequency-domain radar representations remain largely unexplored for Italian Sign Language. In this work, we present the first benchmark for radar-based recognition us ing frequency-domain Range-Time Maps (RTMs) on the MultiMeDaLIS dataset. We define a reproducible prepro cessing pipeline that converts the three synchronized receive channels of a 60GHz radar sensor into compact RTMs, reducing storage requirements and processing time com pared with the original time-domain signal stream. To pro vide a strong baseline for this setting, we introduce SABRE (Sign Attention-based Bidirectional Recurrent Encoder), a model that combines a pretrained backbone with bidirec tional temporal modeling and attention-based pooling. Ex perimental results show that SABRE achieves 81.6% accu racy, outperforming all established baselines. These find ings position this work as the first systematic evidence of frequency-domain RTMs for Italian Sign Languagerecogni tion, suggesting their promise for future research on radar based sign recognition.
A benchmark for radar-based Italian sign language recognition using frequency-domain range-time maps
Raffaele Mineo
;Amelia Sorrenti;Gaia Caligiore;Federica Proietto Salanitri;Giovanni Bellitto;Sabina Fontana;Egidio Ragonese;Concetto Spampinato;Simone Palazzo
2026-01-01
Abstract
Radar has emerged as a promising sensing modality for sign language recognition in privacy-sensitive scenarios, yet frequency-domain radar representations remain largely unexplored for Italian Sign Language. In this work, we present the first benchmark for radar-based recognition us ing frequency-domain Range-Time Maps (RTMs) on the MultiMeDaLIS dataset. We define a reproducible prepro cessing pipeline that converts the three synchronized receive channels of a 60GHz radar sensor into compact RTMs, reducing storage requirements and processing time com pared with the original time-domain signal stream. To pro vide a strong baseline for this setting, we introduce SABRE (Sign Attention-based Bidirectional Recurrent Encoder), a model that combines a pretrained backbone with bidirec tional temporal modeling and attention-based pooling. Ex perimental results show that SABRE achieves 81.6% accu racy, outperforming all established baselines. These find ings position this work as the first systematic evidence of frequency-domain RTMs for Italian Sign Languagerecogni tion, suggesting their promise for future research on radar based sign recognition.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


