ETISEO, performance evaluation for video surveillance systems

Abstract

Abstract

Video Surveillance Online Repository (ViSOR): an integrated framework

“…An example is ontology as defined in the Etiseo project [12] or the result of the "Challenge Project on Video Event Taxonomy" sponsored by the Advanced Research and Development Activity (ARDA) [26]. In [27], a Video Event Representation Language (VERL) is presented which describes an event ontology, associated with Video Event Markup Language (VEML) for event instance annotation.…”

“…Each video and annotation are cataloged together with a set of metadata, such as author, creation date and description. The ontology described in section 4 is also stored in the database by means of three tables: the "Descriptor" by ViSOR is ViPER XML [8], developed at the University of Maryland, since it satisfies several requirements: it is flexible, the list of concepts is customizable and it is widespread (e.g., it is used by Pets [9] and Etiseo [12]). Kasturi et al in their very recent work [29] on performance evaluation adopted the ViPER format as well.…”

“…Some open source tools, such as concepts and events in videos, generate ground truth and annotate videos into XML files. The ViPER annotation format [8] is widely exploited by available vide databases, which are created in workshop and conference contests, like the PETS workshop series [9] or the VSSN workshops of the ACM Multimedia Conference [10] and in those environments that become available from some European or national projects such as I-Lids [11] and Etiseo [12]. These contests propose their own datasets.…”

Tracking people under heavy occlusions by layered data association

“…The third video set is called ETISEO [13]. Amongst all the videos, we selected the video recorded in the central hall of a metro station for its challenging occlusions (see Fig.…”

“…Instead, modelling the matching case is possible by identifying the features' expected range of variations over single targets. To this aim, we conducted a major preliminary experiment by manually annotating the target and its parts in each frame of several videos from various datasets [13,15,1]. In detail, we have manually annotated 20 video sequences over different scenarios such as a corridor, a train station, a shopping mall, and various outdoor scenes, each comprising between 1,000 and 4,000 frames [13, 15, 1] for a total of over 54,000 frames.…”