Sign Language/Gesture Recognition on OOD Target Domains Using UWB Radar

Radar sensors have the advantages of penetrating obstacles to receive echo signals, detecting long distances, being sensitive to micromovements, being compact, lightweight, and portable, and working normally in a variety of lighting and weather situations. More importantly, the radar sensor’s echo signal does not contain any image or video data, thereby protecting the privacy of users and their personal information security. Radar-based gesture/sign language (SL) recognition technology can enable machines to comprehend the meaning of human hand movements and can be applied in a variety of fields, including human–computer interaction, information exchange, home care, and so on, which has significant research value and broad social benefits. Currently, recognition models driven by machine learning require prebuilt datasets. Therefore, the first stage is to design experiments for radar echo signal acquisition for gesture/SL motions in order to obtain sufficient sample data for driving model training. However, the endeavor to capture radar echo signals can only be conducted under specific environments and operating settings. When a gesture/SL recognition model is applied to never-before-seen application scenarios and operating settings, using sample data gathered in specific application scenarios and operating settings may impair generalization performance. Methods for increasing the model’s generalization performance under varied scenarios and operating settings must, thus, be researched, with the goal of enabling the recognition model to have strong generalization performance for scenes or operating settings that were not seen during training.