Please use this identifier to cite or link to this item: https://hdl.handle.net/10356/171887
Title: Human action recognition in the dark based on transferring learning and LSTM
Authors: Yang, Fan
Keywords: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
Issue Date: 2023
Publisher: Nanyang Technological University
Source: Yang, F. (2023). Human action recognition in the dark based on transferring learning and LSTM. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/171887
Abstract: Today, information technology is pervasive worldwide. Throughout the years, it has undergone continuous transformation and will continue to evolve. From the initial invention of the computer to the development of software and artificial intelligence, a new era in human life has been unveiled. Based on deep learning technology, a modern concept called “ smart city ” is proposed. A smart city is a municipality that uses information and communication technologies(ICT) to increase operational efficiency, share information with the public and improve both the quality of government services and citizen welfare. This study introduce an innovative method for conducting video classification in low-light conditions, specifically targeting Human Action Recognition. The majority of prior research has primarily concentrated on datasets characterized by ample brightness levels, facilitating ideal classification conditions. In contrast, the approach revolves around a series of hybrid classifiers, comprising both feature extraction and classification components. The deep learning network of this study, customized for Human Action Recognition in low-light environments, is founded on transfer learning and LSTM (Long Short-Term Memory), effectively balancing accuracy and execution speed requirements. This project encompasses the creation and deployment of a set of hybrid networks composed of diverse components, followed by a performance comparison. These networks harness classic deep learning architectures to extract essential feature vectors, contributing to the advancement of 'smart city' initiatives, ultimately enhancing social security while reducing the expenses associated with public security maintenance. The algorithm is implemented in MATLAB, enabling efficient and reliable computations. Subsequently, data analysis was carried out to assess the comparative effectiveness of various methodologies. The results revealed that the choice of logic significantly impacted accuracy and speed, resulting in superior levels of both. In the realm of classification networks, the author opted for BiLSTM (Bidirectional Long Short-Term Memory) instead of the standard LSTM. BiLSTM processes input sequences bidirectionally, harnessing information from both ends of the sequence, which contributes to enhanced accuracy and efficiency. The author conducted a thorough analysis of the achieved results and offered potential explanations. Additionally, this study have provided recommendations for future research endeavors aimed at addressing areas for improvement.
URI: https://hdl.handle.net/10356/171887
Schools: School of Electrical and Electronic Engineering 
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:EEE Theses

Files in This Item:
File Description SizeFormat 
dissertation_final_version_with_signature.pdf
  Restricted Access
2.79 MBAdobe PDFView/Open

Page view(s)

137
Updated on Jun 25, 2024

Download(s)

22
Updated on Jun 25, 2024

Google ScholarTM

Check

Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.