Chapter 14 - A study on smartphone sensor-based Human Activity Recognition using deep learning approaches

Abstract

With the advent of smart devices like smartphones and smartwatches, it is becoming very easy to track our daily activities using the data captured by the sensors present in these devices. Human Activity Recognition (HAR) is a rapidly growing research field in the domain of computer vision, where a sequence of data for a specified time span is collected from the sensors like accelerometer and gyroscope present in these smart devices. HAR is not only limited to sensor-based observation but also helpful in detecting human activities from still images and videos. HAR plays a crucial role in detecting a user interaction with the environment, which helps in surveillance, health care, and building a smart environment based on human-computer interaction. In this work, a comparative study of the performances of various deep learning–based classification models applied to HAR has been presented. This study also reveals that the deep learning models described in our study provide many insights to predict human activities from sensor data. It is to be noted that designing a good set of handcrafted features from this vast amount of raw sensor data is challenging, thereby building a near-perfect predictive model from the heuristic of those features becomes difficult also. Hence, the use of deep learning–based models for this purpose is a viable alternative. To show the usefulness of deep learning models in handling HAR-related issues, in this work, an exhaustive comparative study of five existing deep learning models which include Convolutional Neural Network (1D-CNN), Recurrent Neural Network with Long-Short Term Memory (RNN-LSTM), CNN-LSTM, ConvLSTM, and Stacked-CNN is performed. Here, these deep-learning models can automatically infer meaningful information from raw sensor data. Three benchmark HAR datasets, namely RealWorld HAR, MHEALTH (Mobile HEALTH), and UCI HAR datasets have been used to demonstrate the task of activity recognition collected from smartphone sensor data. These models have achieved promising results on all the datasets mentioned above.