Visualization of Important Human Motion Feature Using Convolutional Neural Network
Masashi Fukui, Genki Kokubun, Takahiro Nozaki · 2020
Human motion feature extraction is necessary for robot motion generation. In particular, feature extraction methods related to non-periodic motion should be proposed. Recently, the number of the human motion recognition studies utilizing Convolutional Neural Network (CNN) is increasing due to the brilliant ability of extracting and identifying features. CNN has the same or better discrimination ability than humans so that the features extracted by CNN are thought to be useful for human motion understanding. However, since the internal structure of CNN is like a black box, it is difficult to understand the extracted features. In order to understand the features, some visualization methods which are utilized in the image field are applied in this paper. Furthermore, the many conventional studies utilizing the gradient are not preferred for visualizing CNN which recognizes human motion because human motion is sensitive for the object which treated. That is, the gradient method does not work because the value varies greatly depending on the environment even in the same motion. Therefore, new visualizing method without using the gradient is proposed in this paper. The proposed method visualizes CNN focusing part by following the neuron in CNN. Since this method does not require the gradient, the stable and accurate visualization can be performed. The effectiveness is shown by experiments.