Mediapipe-based keypoint extraction with optimized neural models for offline smart control

Hand gesture recognition is a natural way of interaction between humans and computers. Among the many areas where it could be applied, smart homes and assistive systems are the most interesting ones. Still, most methods currently in use require sophisticated systems and cloud computing, thus, the setup causes latency, real-time and battery-operated applications are thereby limited. The present work proposes a straightforward, keypoint-based gesture recognition framework that employs the MediaPipe library for the efficient extraction of landmarks and optimized neural network classifiers for decision-making. By concentrating on four main gestures—Palm (ON), Fist (OFF), Thumbs Up (Increase), and Thumbs Down (Decrease)—the system enables offline, reliable, and fast control of home appliances. The experimental results have proved that the method reaches high accuracy while maintaining low computational cost, hence it becomes a suitable technology for embedded and real-time applications.