Human interactive behaviour recognition method based on multi-feature fusion
Qing Ye, Rui Li, Hang Yang, Xinran Guo · International Journal of Computational Science and Engineering · 2022
Recently, the selection of the overall and individual characteristics in interactive actions and the high-dimensional complexity of features are still important factors affecting the recognition accuracy. In this paper, we propose a human interactive behaviour recognition method based on multi-feature fusion, which includes two parts, feature extraction and behaviour recognition. Firstly, we use histogram feature descriptors to form a three-dimensional gradient histogram of local space-time feature (3D-HOG) and histogram of global optical flow feature (HOF). Then the bag-of-words model is used to reduce the dimensions and the classification matrix is obtained through multilayer perceptron (MLP) classifiers. In the second part, we use recurrent neural network (RNN) to get connections in time. Considering the information of interactive behaviour will be different at different stages, an improved Gauss neural network are proposed for interactive behaviour recognition. The experimental results show that the algorithm can effectively improve the accuracy in the UT-Interaction dataset.