A rehabilitation activity monitoring method based on Shallow-CNN

Abstract

This paper proposes a shallow convolutional neural network (CNN) model to improve the efficiency and accuracy of real-time human activity recognition (HAR). In the traditional convolutional network, an Mix-Patch-Layer (MPL) block based on the attention mechanism is added to enhance the expressiveness of the network extracted features. This block makes the features in the network focus on the information between different parts of itself, which makes up for the loss of global information in temporal data features. Experiments show that the block can improve real-time human recognition accuracy and efficiency with a shallow network.