Skip navigation


  • DSpace logo
  • Home
  • Collections
  • Researcher Profile
  • Explore by
    • Researcher Profile
  • VGU library
  • Help
  • User Guide
  • Sign on to:
    • My DSpace
    • Receive email
      updates
    • Edit Account details

VGU RESEARCH REPOSITORY


Please use this identifier to cite or link to this item: https://epub.vgu.edu.vn/handle/dlibvgu/1728
DC FieldValueLanguage
dc.contributor.authorPham Viet Hoangen_US
dc.date.accessioned2023-10-26T06:25:59Z-
dc.date.available2023-10-26T06:25:59Z-
dc.date.issued2018-
dc.identifier.urihttps://epub.vgu.edu.vn/handle/dlibvgu/1728-
dc.description.abstractThe current concept of Smart Cities aims to provide modern, secure, and sustainable infrastructure and give residents a decent quality of life. To achieve this goal, video surveillance cameras have been deployed to enhance the safety and well-being of the citizens. However, abnormal event detection in surveillance video systems is challenging and requires exhaustive human efforts despite technical developments in modern science. In this project, end-to-end crime-scene anomaly detection on surveillance camera systems is proposed and implemented directly on edge devices. Although the output only determines the scene to be normal or abnormal, the concept of anomaly events stands for nine types of crimes scene: Arson, Assault, Burglary, explosion, Fighting, Road Accident, Shooting, Stealing, and Vandalism. The project is a deep learning-based method with the help of some edge AI techniques and code-based structure from the Facebook open-source Slowfast project, which is re-implemented to apply to the AI system successfully. The system collects camera video data and performs anomaly detection directly on Jetson edge devices with the highest inference speed. I searched for the state-of-the-art (SOTA) models based on the UCFCrime dataset for anomaly detection problems on surveillance videos problem on Paperwithcode and chose weakly-supervised video anomaly detection with robust temporal feature magnitude learning (RTFM) as the final anomaly detection AI solution. The metric for the evaluation process is Area Under the ROC curve (AUC). The AI solutions achieved a 0.8439 AUC score on the UCF-Crime dataset and a 0.884 AUC score on the VNAnomaly dataset collected in Vietnam from Ho Chi Minh City University of Information Technology (UIT). The system was also successfully deployed and tested on Jetson Nano with 1.55 FPS inference speed by using TensorRT edge AI techniques with only 2.61 GB RAM usage total.en_US
dc.language.isoenen_US
dc.publisherVietnamese-German University-
dc.rightsAttribution-NonCommercial 4.0 International*
dc.rights.urihttps://creativecommons.org/licenses/by-nc/4.0/*
dc.subjectAnomaly detectionen_US
dc.subjectJetson Edge computingen_US
dc.subjectEdge AIen_US
dc.subjectSlowfasten_US
dc.titleCrime-scenes detection from surveillance videos directly on Jetson edge devices systemen_US
dc.typeThesisen_US
item.fulltextWith Fulltext-
item.languageiso639-1other-
item.grantfulltextopen-
Appears in Collections:Computer Science (CS)
Files in This Item:
File Description SizeFormat
Crime scenes detection from surveillance videos directly on Jetson edge devices system.pdf12.65 MBAdobe PDFView/Open
Show simple item record

Page view(s)

30
checked on Oct 16, 2025

Download(s)

10
checked on Oct 16, 2025

Google ScholarTM

Check


This item is licensed under a Creative Commons License Creative Commons

© Copyright 2020 by Vietnamese - German University Library.
Add: Ring road 4, Quarter 4, Thoi Hoa Ward, Ben Cat City, Binh Duong Province
Tel.:(0274) 222 0990. Ext.: 70206