Comparative Analysis of Lightweight Gesture Recognition Based on Raspberry Pi Devices
TL;DR
Experimental results show that MobileNetV2 outperforms ResNet50 in both accuracy and inference speed in the original state, and quantization technology can significantly improve its deployment performance.
Abstract
With the rapid development of edge computing and the Internet of Things. Deploying deep learning models on resource-constrained devices such as Raspberry Pi has become a research hotspot. Gesture recognition is a natural human-computer interaction method. It has broad application prospects in fields such as smart homes and virtual reality. However, balancing real-time performance and accuracy on edge devices such as Raspberry Pi remains a challenge. This paper uses MobileNetV2 and ResNet50 models as research methods. By performing post-training integer quantization processing, an analysis of the balance between model training performance and speed on edge devices such as Raspberry Pi. Experimental results show that MobileNetV2 outperforms ResNet50 in both accuracy and inference speed in the original state. After quantization, the accuracy of MobileNetV2 improved to 94.62%. The inference speed has been increased to 39.63 FPS, meeting the requirements for real-time recognition. After ResNet50 quantization, the accuracy dropped to 63.17%, while the speed only increased to 4.53 FPS. In summary, MobileNetV2 has a greater advantage on edge devices, and quantization technology can significantly improve its deployment performance. This article provides a reference for selecting gesture recognition models for edge devices such as Raspberry Pi.