In this project, we explore a novel real-time gesture recognition system that based on a Sony IMX500 AI camera on Raspberry Pi. Traditional approaches to gesture recognition typically rely on high-performance GPUs or cloud-based computation, resulting in higher latency, higher cost, and reduced portability. To address these challenges, this work implements gesture recognition directly using the on-module AI processor, resulting in significantly improved computational efficiency, portability, and responsiveness. We developed a highly optimised AI model using BrainBuilder software, which was trained using a dataset of real captured images and synthetic augmented data generated through advanced 3D modelling techniques.
Experimental results show that the model performs excellently, with gesture classification accuracy and recall consistently above 95%. Furthermore, the system demonstrates robust performance across a wide range of environmental conditions, with extremely low latency and efficient real-time processing power, suitable for practical interactive applications such as the rock-paper-scissors game demonstrated. This work provides practical and innovative solutions for low-cost, efficient and robust hand gesture recognition for commercial applications in interactive entertainment, healthcare and smart home systems. Future improvements include expanding the diversity of the data set, further optimising the neural network architecture, and exploring a wider range of applications and commercial potential, while considering the ethical and privacy issues associated with the processing of biometric data.
Comments