Date of Award
Summer 8-2021
Degree Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
School
Computing Sciences and Computer Engineering
Committee Chair
Zhaoxian Zhou
Committee Chair School
Computing Sciences and Computer Engineering
Committee Member 2
Chaoyang Zhang
Committee Member 2 School
Computing Sciences and Computer Engineering
Committee Member 3
sarah Lee
Committee Member 3 School
Computing Sciences and Computer Engineering
Committee Member 4
Zhanxin Sha
Committee Member 4 School
Kinesiology and Nutrition
Committee Member 5
Ras B Pandey
Committee Member 5 School
Mathematics and Natural Sciences
Abstract
Human Activity recognition, with wide application in fields like video surveillance, sports, human interaction, elderly care has shown great influence in upbringing the standard of life of people. With the constant development of new architecture, models, and an increase in the computational capability of the system, the adoption of machine learning and deep learning for activity recognition has shown great improvement with high performance in recent years. My research goal in this thesis is to design and compare machine learning and deep learning models for activity recognition through videos collected from different media in the field of sports.
Human activity recognition (HAR) mostly is to recognize the action performed by a human through the data collected from different sources automatically. Based on the literature review, most data collected for analysis is based on time series data collected through different sensors and video-based data collected through the camera. So firstly, our research analyzes and compare different machine learning and deep learning architecture with sensor-based data collected from an accelerometer of a smartphone place at different position of the human body. Without any hand-crafted feature extraction methods, we found that deep learning architecture outperforms most of the machine learning architecture and the use of multiple sensors has higher accuracy than a dataset collected from a single sensor.
Secondly, as collecting data from sensors in real-time is not feasible in all the fields such as sports, we study the activity recognition by using the video dataset. For this, we used two state-of-the-art deep learning architectures previously trained on the big, annotated dataset using transfer learning methods for activity recognition in three different sports-related publicly available datasets.
Extending the study to the different activities performed on a single sport, and to avoid the current trend of using special cameras and expensive set up around the court for data collection, we developed our video dataset using sports coverage of basketball games broadcasted through broadcasting media. The detailed analysis and experiments based on different criteria such as range of shots taken, scoring activities is presented for 8 different activities using state-of-art deep learning architecture for video classification.
Copyright
Shakya, 2021
Recommended Citation
Shakya, Sarbagya, "Vision Based Activity Recognition Using Machine Learning and Deep Learning Architecture" (2021). Dissertations. 1926.
https://aquila.usm.edu/dissertations/1926