There has been a great effort to use technology to make exercise more interactive, measurable and gamified. However, in order to optimize the detection accuracy, these efforts have always translated themselves into motion detection with multiple sensors including purpose specific hardware, which results in extra expenses on both the content production and consumption and induces limitations on the final mobility of the user. In this paper we aim to improve the accuracy, learning speed and detail range of Posenet’s AI sensorless human pose detection by using an artificial neural network to optimize its extraction and comparison algorithms, changing the current model that uses a ResNet convolutional neural network (CNN) to a model using DenseNet and developing a new algorithm for detailed corrections using relevant artificial neural networks. The findings here will be applied on a posture correction system for a dance and fitness application.

Keywords : Sensorless human pose estimation – Artificial Intelligence (A.I.) – Machine learning – Posenet – DenseNet – Posture correction methods – Human motricity – Motion capture methods – Dance – K-pop – E-sports – South Korea.