Abandoned object detection and classification

Abstract

Video surveillance in public places has witnessed tremendous growth over the last decade with applications based on closed-circuit security systems to IP cameras. Among the various aspects of surveillance, automatic detection and classification of abandoned objects could serve as a valuable application reducing potential threats to public safety. The aim of this project is to detect the unattended abandoned objects from videos of public places and classify the detected objects into one of the pre-defined object classes. First a background model of the scene under consideration is constructed. Background subtraction is then carried out on k frames to generate k foreground masks which are intersected to produce static foreground objects which are likely to be either abandoned objects or static humans. The detected foreground regions are then classified into different categories of interest. Foreground regions identified as humans are discarded while for other static and potentially suspicious objects, the owner of the object is sought within a predefined neighborhood of the detected item. Presence of owner within the neighborhood leads to the assumption that the object is attended. In case the owner is not found, the system back tracks the video to the point while the object was still attended and tracks the owner from the point until the owner is no more in the camera view resulting in generation of an alarm. The system is evaluated on a set of five video sequences from two standard data sets, PETS 2006 & 2007 and i-LIDS. The scenarios considered assume that each item of luggage has one owner and each person owns at most one item of luggage. The system successfully identified all cases where the individuals leave the scene without their luggage. The tracking module however has limitations that it is not able to follow the owner in case of occlusion which may result in false alarms. The detection methodology being based on an unsupervised approach does not require prior training of objects to be detected, and hence successfully detects objects of all shapes, sizes and orientations. The proposed system can be extended to more complex scenarios which are true representatives of the real world situations. A multi camera network may also be considered in this regard providing information from different perspectives.