Human Activity Recognition and Temporal Action Localization Based on Depth Sensor Skeletal Data Iskelet Verisine Dayali Insan Aktivitesi Tanima ve Zamanda Aktivite Yeri Belirleme

Gorgulu Y. E., TAŞDELEN K.

2020 Innovations in Intelligent Systems and Applications Conference, ASYU 2020, İstanbul, Turkey, 15 - 17 October 2020, (Full Text) identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/asyu50717.2020.9259886
  • City: İstanbul
  • Country: Turkey
  • Keywords: Deep learning, Human activity recognition, LSTM, RNN
  • Isparta University of Applied Sciences Affiliated: No

Abstract

Abundance in inertia measurement units and wide usage of depth sensors has led to the effortless acquisition of significant activity data. As a result, studies in the area of human activity recognition and datasets related to this area have risen. In this paper, skeletonized action sequence data; obtained by a depth sensor, is used to classify among twenty seven different class of activities. A three-layer Long-Short Term Memory(LSTM) architecture was used to analyze the mentioned skeleton data. LSTM cells are a special type of recurrent neural network which can remember long time relationship between sequences of related data. Also, another application was carried out using a recurrent neural network(RNN) architecture to address where a specific action occurs on a video stream. This architecture consists of four layers. Thanks to this deep learning structures; which led to the automated extraction of features, high classification accuracy rates around %93 can be achieved.