Designing an AI-assisted toolbox for fitness activity recognition based on deep CNN

Online applications for virtual meetings have be-come handy for many fitness instructors these days. However, currently used applications have limitations in tracking clients' performance simultaneously, particularly for many participants. In this paper, we proposed a fitness AI-assisted toolbox to detect the fitness activities of athletes from steamed frames. The innovative Magnitude flow method is introduced to compute the dense optical flow of each frame and provides details of human body poses during exercising. The two-stream convolutional neural network is proposed to only detect fitness actions from RGB frames and improved magnitude flows. The two-stream CNN is designed with a MobileNetVl backbone and a Multi-layer perceptron (MLP) network as the classification head. Moreover, we introduced an Observer network as a score fusion approach to improve the final detection accuracy and reduce the uncertainties between the two networks. Our proposed method is tested over selected fitness activities of the UCF dataset. The plenary results were acquired by testing selected activities and considering their ability in detecting major fitness actions. The efficient accuracies and confusion matrices describe that our proposed model is an effective solution to assist fitness instructors in tracking, analyzing, and conveniently supervising many participants.