Hand Gesture Recognition for Traffic Signals Based on Improved Lightweight Convolutional Neural Network

Gesture recognition is a crucial aspect of the development of intelligent transportation systems. In this paper, we propose a novel method for recognizing gestures of artificial traffic signals on embedded devices using a lightweight convolutional neural network (CNN). Our approach adopts GhostNetV2 as the base model and improves it by incorporating the MobileNeXt Hourglass block, resulting in GhostNetV2-A, which effectively reduces the size of the convolutional layers and enhances the device's operational capability under limited conditions. In terms of hardware, we replaced the original SPI flash with a 32GB storage card, used a smaller display screen, and added a user button to address the limitations of OpenMV4 Plus and improve its usability. Finally, we deployed the model on the device, and its performance was evaluated in a self-built experimental environment, achieving an average accuracy of 81.3%. The proposed method achieves an average recognition rate of about 10FPS on improved embedded devices. This method provides a practical and efficient solution for real-time gesture recognition of artificial traffic signals on embedded devices, with potential applications in various traffic scenarios.