A hybrid framework combining background subtraction and deep neural networks for rapid person detection

Chulyeon Kim, Jiyoung Lee, Taekjin Han, Young Min Kim

Research output: Contribution to journalArticlepeer-review

19 Scopus citations


Currently, the number of surveillance cameras is rapidly increasing responding to security issues. But constructing an intelligent detection system is not easy because it needs high computing performance. This study aims to construct a real-world video surveillance system that can effectively detect moving person using limited resources. To this end, we propose a simple framework to detect and recognize moving objects using outdoor CCTV video footages by combining background subtraction and Convolutional Neural Networks (CNNs). A background subtraction algorithm is first applied to each video frame to find the regions of interest (ROIs). A CNN classification is then carried out to classify the obtained ROIs into one of the predefined classes. Our approach much reduces the computation complexity in comparison to other object detection algorithms. For the experiments, new datasets are constructed by filming alleys and playgrounds, places where crimes are likely to occur. Different image sizes and experimental settings are tested to construct the best classifier for detecting people. The best classification accuracy of 0.85 was obtained for a test set from the same camera with training set and 0.82 with different cameras.

Original languageEnglish
Article number22
JournalJournal of Big Data
Issue number1
StatePublished - 2018 Dec 1


  • Background subtraction
  • CCTV
  • Convolutional Neural Network
  • Object detection


Dive into the research topics of 'A hybrid framework combining background subtraction and deep neural networks for rapid person detection'. Together they form a unique fingerprint.

Cite this