The Variegated Applications of Deep Learning Techniques in Human Activity Recognition

Abstract

Humans are vivacious and obstinate in that they are plagued by a constant need to be motile and sprightly which gives us a goldmine of data to work on. This constant stream of ever-changing activities performed by us could be dissected fastidiously to gain insight into the specifications of that activity. This could spur the use of IoT, automated devices and real-time monitoring. A variety of techniques some of which are video camera feeds which could be sourced from CCTVS and sensors, could be put to use to efficaciously procure data. This paper will delve into the various techniques proposed by various researchers and compare their performance on various deep learning and machine learning models to analyse them intrinsically. We will also showcase our own model consisting of the use of a 3D tempo-spatial dataset called the UCI-HAR dataset employing various deep learning models like LSTM, SVMs and more. The deep learning model will be improved upon by architectural and hyper parameter improvements. Other sections will discuss the related works including the datasets used in Human Activity Recognition. Also contained in the discussion section are the technicalities of the papers like the accuracy and the relevancy of the deep learning models being used. A proposed hybrid models using both video feed and sensor data for recognition will be floated. A panoply of industries including the health and defence sectors stand to gain from the rapid recognition of human activities.