A Lightweight CNN for RGB-D Grasp Classification in Computer Vision-Based Prosthetic Hands

Prosthetic hand development has progressed significantly from basic mechanical designs to advanced systems that incorporate computer vision for enhanced functionality. However, the challenge now is creating a lightweight, accurate and computationally efficient model suitable for integrating embedded microcontrollers in prosthetic devices. This paper presents a novel framework based on a Convolutional Neural Network (CNN) model with just 160,000 trainable parameters and a size of under 1MB, making it viable for deployment in microcontrollers like Arduino, ESP, and Raspberry Pi. The framework has a comprehensive input of 32x32x3 which incorporates grey-scaled RGB, depth, and masked data. The presented model is tested on a newly structured RGB-D dataset containing 51 objects across 7 grasp types, achieving 99% accuracy with minimal computational resources.