Lightweight Human Behavior Recognition Method for Visual Communication AGV Based on CNN-LSTM

Abstract

Behavior recognition uses deep learning network model to automatically extract the deep features of data, but traditional machine learning algorithms have some problems such as manual feature extraction and poor generalization ability of models. The S-MobileNet is proposed for human behavior recognition. Firstly, the 3D convolution to extract features is used to build a time series model to learn the long-term dependence of human behavior characteristics on time series. Secondly, Long Short-Term Memory (LSTM) is used as the input of multi-layer recurrent neural network time series model, so as to obtain individual dynamic features, and then individual features are aggregated by attention pooling mechanism to obtain corresponding group behavior features. At last, the recognition of individual behavior and group behavior is completed by relying on the characteristics of individual and group behavior. The experiments show that the network in this paper achieves high recognition accuracy on UCF101 and HMDB51 datasets, and the overall recognition rate of proposed model for 13 kinds of human behaviors is 95.3%.