Human Activity Recognition u sing Resnet 34 Model
Akansha Abrol1, Anisha Sharma2, Kritika Karnic3, Raju Ranjan4
1Akansha Abrol*, Department of Computing Science and Engineering, Galgotias University, Greater Noida (U.P), India.
2Anisha Sharma, Department of Computing Science and Engineering, Galgotias University, Greater Noida (U.P), India.
3Kritika Karnic, Department of Computing Science and Engineering, Galgotias University, Greater Noida (U.P), India.
4Raju Ranjan Department of Computing Science and Engineering, Galgotias University, Greater Noida (U.P), India.
Manuscript received on March 11, 2021. | Revised Manuscript received on April 30, 2021. | Manuscript published on May 30, 2021. | PP: 224-228 | Volume-10 Issue-1, May 2021. | Retrieval Number: 100.1/ijrte.A58960510121 | DOI: 10.35940/ijrte.A5896.0510121
Open Access | Ethics and Policies | Cite | Mendeley
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Activity recognition has been an emerging field of research since the past few decades. Humans have the ability to recognize activities from a number of observations in their surroundings. These observations are used in several areas like video surveillance, health sectors, gesture detection, energy conservation, fall detection systems and many more. Sensor based approaches like accelerometer, gyroscope, etc., have been discussed with its advantages and disadvantages. There are different ways of using sensors in a smartly controlled environment. A step-by-step procedure is followed in this paper to build a human activity recognizer. A general architecture of the Resnet model is explained first along with a description of its workflow. Convolutional neural network which is capable of classifying different activities is trained using the kinetic dataset which includes more than 400 classes of activities. The videos last around tenth of a second. The Resnet-34 model is used for image classification of convolutional neural networks and it provides shortcut connections which resolves the problem of vanishing gradient. The model is trained and tested successfully giving a satisfactory result by recognizing over 400 human actions. Finally, some open problems are presented which should be addressed in future research.
Keywords: Video Surveillance, Resnet, Convolutional Neural Network, Kinetic Dataset.