Real-Time Hand Gesture Recognition: Lightweight Keypoint-Based Approach with Medoid Similarity

Abstract

Gesture recognition is crucial in computer vision, with applications in security, consumer electronics, and beyond. While deep learning techniques like Convolutional Neural Networks or pre-processing methods such as optical flow achieve high classification accuracy, they often rely on large pre-trained networks or computationally intensive preprocessing, making them unsuitable for real-time or resource-limited applications. This paper introduces a novel lightweight classifier that leverages skeleton features and hand-shape similarity to representative gestures for efficient gesture recognition. By focusing on keypoints that reflect the human body’s structure and similarity to static gesture medoids, our approach reduces computational complexity while maintaining competitive performance, as demonstrated in tests on the public Jester database. This work offers an efficient solution suitable for gesture recognition applications requiring real-time processing with limited computational resources.