Lightweight Hand Gesture Recognition Using MediaPipe Landmark Features and LSTM for Embedded Human–Machine Interaction

Abstract

Hand gesture recognition has emerged as an effective approach for enabling natural and contactless human–machine interaction (HMI). However, many existing vision-based approaches rely on computationally intensive image-based deep learning models or require specialized sensing hardware, limiting their deployment on low-cost real-time systems. This paper presents a lightweight vision-based hand gesture recognition framework that integrates MediaPipe hand landmark detection, geometric feature extraction, long short-term memory (LSTM) classification, and Arduino-based device control. Instead of directly processing RGB images, the proposed approach represents hand gestures using a compact feature vector consisting of normalized landmark coordinates, inter-landmark distances, and joint angles. Temporal dependencies between consecutive hand movements are modeled using an LSTM network to recognize both static and dynamic gestures. A prototype system was developed using a standard RGB webcam and an Arduino microcontroller for real-time control of electronic devices and environmental monitoring. Experimental results demonstrate an overall recognition accuracy of 92.5% with inference latency below 100 ms. The model contains only 24.6K parameters, occupies 0.3 MB of storage, and achieves 12 ms inference time on a standard CPU, enabling responsive interaction suitable for real-time HMI applications. The proposed framework provides a low-cost and computationally efficient solution for gesture-based control systems in smart environments, assistive technologies, and embedded applications.