Efficient Gesture Recognition Using Simplified and Quantized CNN Architecture

Abstract

In this paper, we propose an efficient technique through simplification and quantization of the structure of the conventional convolutional neural network (CNN) based model for real-time processing and low-power implementation of UWB radar-based gesture recognition system. Recently, gesture recognition technology has been used in the fields of human-computer interaction (HCI) and smart device control, especially contactless recognition methods using UWB radars are advantageous for privacy protection. Conventional CNN-based gesture recognition models achieved high accuracy, but it was difficult to apply edge devices in terms of model size and amount of computation. In this paper, to overcome this limitation, the number of hidden layers is limited to three and the model is effectively lightened by applying 8-bit quantization based on post-processing. As a result of the experiment, the accuracy of the proposed model recorded 96.87%, which is similar to that of the existing CNN, and the efficiency is confirmed through a significant weight reduction effect in the amount of computation and model size. The proposed model is suitable for a real-time gesture recognition system in edge device and embedded environments.