Radar-based Gesture Classification Using a Variational Auto-Encoder Neural Network Algorithm

Simple SummaryContent extracted from patent full text and abstract with AI.

This invention relates to a method and system for classifying human gestures using radar sensors and a variational autoencoder neural network (VAENN). By analyzing radar measurements (like hand or finger motions), and processing them into positional time spectrograms, the system can accurately identify gesture types (e.g., swipes, circles, finger movements) in a robust and privacy-protecting way. The neural network is trained using advanced loss functions (including triplet and center loss with statistical distances) to enhance discrimination between gesture classes and minimize misclassification, even in noisy or varied environments.

Use CasesContent extracted from patent full text and abstract with AI.

• Touchless control of smartphones, tablets, or computers through hand gestures.
• Gesture-based control for smart home devices or appliances (e.g., lights, TV, thermostats).
• Automotive infotainment or vehicle controls via gestures, improving driver safety.
• Enhanced human-machine interaction in AR/VR headsets and environments.
• Public kiosks (e.g., ticketing or vending machines) with gesture control for hygiene and accessibility.
• Wearables and healthcare devices that interpret user gestures for commands or monitoring.
• Gaming interfaces using subtle gesture recognition for improved user experience.

BenefitsContent extracted from patent full text and abstract with AI.

• Operates reliably regardless of lighting conditions or visual obstructions, unlike cameras.
• Protects user privacy as it does not rely on capturing images or video.
• Highly accurate and robust against noise, user variability, and background motion.
• Can reject unknown or random gestures, reducing false positives.
• Works in cluttered or dynamic environments, broadening deployment possibilities.
• Low processing and memory requirements allow for easy integration with embedded devices.
• Easily adaptable for new gesture classes without complete retraining.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

A gesture classification based on radar measurements is disclosed. A variational-autoencoder neural network algorithm is employed. The algorithm can be trained using a triplet loss and center loss. A statistical distance can be considered for these losses.