A Shallow Convolution Network Based Contextual Attention for Human Activity Recognition

Abstract

Human activity recognition (HAR) is increasingly important in ubiquitous computing applications. Recently, attention mechanism are extensively used in sensor-based HAR tasks, which is capable of focusing the neural network on different parts of the time series data. Among attention-based methods, the self-attention mechanism performs well in the HAR field, which establish the correlation of key-query to fuse the local information with global information. But self-attention fails to model the local contextual information between the keys. In this paper, we propose a contextual attention (COA) based HAR method, which utilize the local contextual information between keys to guide learning the global weight matrix. In COA mechanism, we use $k \times k$ kernel to encode input signal to local contextual keys to extract more contextual information between keys. By fusing local key and query to generate global weight matrix, we can establish the correlation between local features and global features. The values are multiplied by the weight matrix to get a global contextual key, which include global contextual information. We combine the local key and global key to enhance feature’s expression ability. Extensive experiments on five public HAR datasets, namely UCI-HAR, PAMAP2, UNIMIB-SHAR, DSADS, and MHEALTH show that the COA-based model is superior to the state-of-the-art methods.