Exploring the role of multi-scale convolutional operators in behavior sequence recognition

Abstract

In order to simultaneously calculate the temporal and spatial characteristics of behavior sequence samples, a convolutional neural network recognition model based on a multi-scale convolutional operator is proposed. Firstly, the skeleton vector information in the sequence samples is integrated into a behavior matrix by superposition, and then the matrix is input into the recognition model. In order to explore the role of bone points with different adjacencies in describing human behavior, the convolutional operator in each layer of the convolutional neural network is extended to a multi-scale convolutional operator, and the features obtained by the network are used for classification. Good recognition rates were obtained in the MSR-Action3D dataset and HDM05 dataset.