Clustering of flag-raising play using IHMM and Convolutional Autoencoder

Abstract

Human motion recognition is a widely known research topic in computer vision. In supervised learning of motion, it is necessary to collect and label a huge amount of data for training. We thought that labeling could be done efficiently by automatically dividing the collected data into several patterns in advance when preparing the teacher data. We extracted the features of the motion data of the flag-raising play using the convolutional autoencoder, and clustered using the Infinite Hidden Markov Model. We use motion capture to measure flag-raising play, and PERCEPTION NEURON is used for motion capture. The flag-raising play consists of raising and lowering the left hand and raising and lowering the right hand. The accuracy of the convolutional autoencoder has been improved by updating the model of the convolutional autoencoder and the preprocessing of the data. we got closer to the number of correct tags by updating the model of the convolutional autoencoder from the result of Infinite Hidden Markov Model.