Human Activity Recognition using A Improved Model Based on Multi-Head CNN-LSTM

Abstract

Recent deep learning methods have been used to obtain good performance for Human Activity Recognition (HAR) task. In this paper, three fully convolutional blocks form Multi-Head Convolutional Neural Network (Multi-Head CNN), which abstracts the high-level features of sensor data. An additional convolutional layer (ACL) is used to make up for the missing low-level features and augmented with each head CNN, which is expressed as a NCNN. Then a Bidirectional-LSTM(BiLSTM) follows each NCNN to form each head CNN+BiLSTM. The output of all three parallel CNN+BiLSTMs is concatenated and conveyed into a dense layer, which is followed by an output layer with six neurons on the basis of the number of output classes. Compared to other models, the experimental results show that the proposed architecture can achieve better performance, which prevent overfitting and get a very good accuracy.