HARDenseNet: A 1D DenseNet Inspired Convolutional Neural Network for Human Activity Recognition with Inertial Sensors
Kiran Mehmood, Hamza Ali Imran, Usama Latif
Abstract
Human Activity Recognition (HAR) is currently one of the active research areas considering its applications in fields such as sports, healthcare, Social interaction, fitness, entertainment and the list goes on. Traditionally Computer Vision (CV) strategies were used for HAR which has many provocations including portability, environmental conditions, occlusion, greater cost and most of all privacy. But recently a variety of sensors such as gyroscope, accelerometer and heart rate sensor etc are used for HAR. There are many benefits of using sensor data as an alternative to usual computer vision techniques. Their usage is said to have removed almost all the limitations of computer vision strategies. The use of Machine Learning (ML) and Deep Neural Networks (DNN) using inertial sensor data for Human Activity Recognition can be extensively found in literature. In this paper, we have proposed a novel 1 dimensional neural network which is inspired by DenseNet neural network which has 1 dimensional convolutional layers for processing 1 dimensional signal data of inertial sensors.