Multidimensional Particle Swarm Optimization Based Long Short-Term Memory Method for Human Activity Recognition

Abstract

Human activity recognition (HAR) is indispensable for human-machine interaction. Related applications in many fields, e.g. medical, military, and security, facilitate and even change the lives of human beings. However, a key issue, i.e. the unsatisfactory recognition accuracy, still remains to be addressed. In this paper, we explore a new way to design a long short-term memory (LSTM) method specifically for HAR with datasets collected by sensors or mobile devices. In particular, combining with the multidimensional particle swarm optimization (MD-PSO), the LSTM method, which has advantages in processing collected datasets with timing characteristics, is not only optimized in terms of parameters (i.e. weights and bias), but also of the network topology as well as the number of blocks in each layer. In other words, we can further get the optimal topology and network mapping parameters of the LSTM network under the conditions of supervised learning. By utilizing the real dataset, performance of the proposed method is verified. Moreover, the test results show that our method significantly outperforms the commonly used support vector machine (SVM) and LSTM methods.