Real-Time Edge AI Sign Language Recognition via Motion History Images and Lightweight CNNs on Embedded Hardware

This is the authors' abstract. We don't add key points for this paper.

Project Title: Real-Time Edge AI Sign Language Recognition on Embedded Systems Hardware Stack: Raspberry Pi 4 Model B, OV5647 Camera Module, 3.5" LCD Core Algorithms: Motion History Image (MHI), Custom 4-Layer CNN (1.8M parameters) Key Metrics: 30 FPS live inference speed, 99.50% offline accuracy (6 gesture classes). Note: The offline split was performed at the frame level; this score does not establish accuracy on new sessions or participants. Live testing showed frame-to-frame prediction instability. Key Focus: Spatial-temporal feature extraction, embedded vision optimization, real-time prediction stability, Pearson correlation diagnostic. This document is the undergraduate graduation thesis (HCMUT, 2021) formatted as a conference paper. Code and dataset available from the authors upon reasonable request.