Title: Detection of abnormal behaviour in surveillance applications
Authors: De Smedt, Floris
Tuytelaars, Tinne
Goedemé, Toon
Issue Date: Mar-2014
Conference: ECUMICT edition:6 location:Ghent, Belgium date:27-28 March 2014
Article number: 30
Abstract: The presence of cameras for surveillance purposes are very common nowadays (train stations, public transport, parking lots, ...). Most of these systems work in a passive way, so the recorded video material can only be used as evidence after the crime. When a real-time alarm is necessary, human intervention (someone watching the video stream during recording) is required. In this PhD-project we develop a system to support the human intervention by indicating the presence of abnormal behavior with a certain probability. Our approach combines computer vision, to analyze the video data, with knowledge representation, to describe, train and detect abnormal behavior. The algorithm we use for object detection is the Deformable Part Model detector proposed by P. Felzenszwalb, which delivers state-of-the-art detection results on a variety of datasets. To meet the real-time requirement of our application, we created a hybrid implementation of this algorithm. Our system performs the calculation of the feature pyramid on GPU (using CUDA) and performs the model evaluation on CPU. By parallellizing these tasks, we obtained a speedup of 12.7x over the vanilla implementation reaching a detection speed of 12.9fps, evaluating all scales and positions. By combining this hybrid implementation with the warping window approach and a tracking-by-detection approach using kalman, we reduced the searching space to a single scale and only a limited amount of positions to evaluate. Hereby we obtained a speed of 500 pedestrian detections per second without the loss of accuracy. Based on the temporal information of our object tracker, we can classify the action of our objects. We will use knowledge representation (KR) to describe the scene we are monitoring. In a first level of complexity, we will use sanity checks. These rules are simple to define. At a second level, we use KR to predict future actions (”entering is followed by paying and sitting down.”). By training the importance of these rules, a probability can be assigned to the abnormality of a certain situation.
Description: De Smedt F., Tuytelaars T., Goedemé T., ''Detection of abnormal behaviour in surveillance applications'', 6th European conference on the use of modern information and communication technologies - ECUMICT 2014, March 27-28, 2014, Gent, Belgium.
Publication status: published
KU Leuven publication type: IMa
Appears in Collections:Electrical Engineering (ESAT) TC, Technology Campus De Nayer Sint-Katelijne-Waver
Technologiecluster ESAT Elektrotechnische Engineering
ESAT - PSI, Processing Speech and Images

Files in This Item:
File Description Status SizeFormat
ecumict2014FDS.pdf Published 36KbAdobe PDFView/Open

These files are only available to some KU Leuven Association staff members


All items in Lirias are protected by copyright, with all rights reserved.