Browsing by Author "Sahle, Daniel"

Now showing 1 - 2 of 2
  • No Thumbnail Available
    Item
    Real-Time Shoplifting Detection From Surveillance Video
    (Addis Ababa University, 2018-10-05) Sahle, Daniel; Assabie, Yaregal (PhD)
    With surveillance cameras being ubiquitous in most big stores while shoplifting breaks owners' banks, better prevention mechanisms for the crime is vital more than ever. This can be achieved through an efficient automatic detection surveillance system that can detect this event. The current methods employed in the industry are not efficient as human operators scanning a lot of screens have their own shortcomings. Human labor is also getting more expensive, especially in urban places where such stores exist in abundance. Existing methods of activity detection do not address the problem as each action has unique characteristics and intricate details that it has to be modeled independently. This thesis introduces a novel hybrid based real-time shoplifting detection architecture that detects the event of shoplifting from surveillance videos. The model, using CNN classification and optical flow features from the sequence of frames, makes use of different features of the event that is learned from video examples and applies different techniques to this to detect when the event occurs in sight of the cameras. Moreover, a joint-based rule-based method of detection of joint proximity is designed. To show the effectiveness of the system, a prototype is developed and tested with a dataset we prepared for the purpose of this specific thesis. The analysis of the evaluation indicates that the system provides an efficient automatic real-time shoplifting detection with 55% recall and 60% precision.
  • No Thumbnail Available
    Item
    Real-Time Shoplifting Detection From Surveillance Video
    (Addis Ababa University, 10/5/2018) Sahle, Daniel; Assabie, Yaregal (PhD)
    With surveillance cameras being ubiquitous in most big stores while shoplifting breaks owners' banks, better prevention mechanisms for the crime is vital more than ever. This can be achieved through an efficient automatic detection surveillance system that can detect this event. The current methods employed in the industry are not efficient as human operators scanning a lot of screens have their own shortcomings. Human labor is also getting more expensive, especially in urban places where such stores exist in abundance. Existing methods of activity detection do not address the problem as each action has unique characteristics and intricate details that it has to be modeled independently. This thesis introduces a novel hybrid based real-time shoplifting detection architecture that detects the event of shoplifting from surveillance videos. The model, using CNN classification and optical flow features from the sequence of frames, makes use of different features of the event that is learned from video examples and applies different techniques to this to detect when the event occurs in sight of the cameras. Moreover, a joint-based rule-based method of detection of joint proximity is designed. To show the effectiveness of the system, a prototype is developed and tested with a dataset we prepared for the purpose of this specific thesis. The analysis of the evaluation indicates that the system provides an efficient automatic real-time shoplifting detection with 55% recall and 60% precision.