Real-Time Classification of Imperfect Gesture Data Utilizing Convolutional Neural Network for Internet of Things (IoT) Applications

This paper deals with the development and application of an Artificial Intelligence (AI) neural network. The network was designed for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{IoT}$</tex> edge devices to make them smarter and more efficient in triaging the input information uploaded to the neural network for further analysis and utilization. It presents the development of an image classification Convolutional Neural Network (CNN) that is optimized for implementation in the new generation of powerful, compact, energy efficient embedded systems being developed for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{IoT}$</tex> edge applications. The network was trained to identify skeleton images in different poses captured using an Xbox 360 Kinect. To accomplish this, various techniques and strategies were employed to train and test the CNN, including the use of various datasets and different configurations of the network's architecture. The results of the study determined the overall accuracy of the CNN to be 99.60%. The results provide valuable insights into the potential of using CNNs for image classification and gesture identification, highlighting the importance of careful design and optimization of the network architecture to achieve optimal performance suitable for <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{IoT}$</tex> edge implementations. Potential applications identified were medical, surveillance, and facial tracking. These are just a few examples of the potential uses of this network for image classification.