Robust detection and tracking of objects (bibtex)
by Florian Seitner
Abstract:
Due to the increasing availability of fast and cheap hardware in the past few years, today a wide range of complex visual tracking tasks is possible. Efficient mathematical methods can provide a high robustness which also makes visual tracking interesting for many industrial purposes. However, the high demands on quality and speed still provide a major challenge for each tracking application. In this thesis a tracking system is introduced, which tries to address both demands appropriately by using currently available algorithms to quickly track pedestrians in video streams. By combining these well-proved algorithms, a good solution regarding computational complexity, accuracy and stability is obtained. To achieve this task, a fast object detector similar to the approach of Viola et al. [Viola 2003] is used as one component in this tracking system. This detector uses Haar-like features which are very fast to compute and makes a quick pedestrian detection in a frame possible. Next to the detection system, an adaptive background model sub-divides each frame into foreground and background regions. As a compromise between complexity and robustness a single-mode parametric background model based on normal distributions and wrapped normal distributions is used. Both background model and detector are combined to provide the tracking system with locations of pedestrian-like regions and to sub-divide the body into three parts: head, upper body and lower body. After this segmentation into finer tracking units a set of colour and spatial features for further tracking is extracted from each part individually. Individual and spatially separated body parts also provide the possibility to use colour histograms in a spatial sense. Moreover, an appearance model provides accurate solutions and approximations when occlusions or missing detections occur.
Reference:
Robust detection and tracking of objects (Florian Seitner), Technical report, PRIP, TU Wien, 2005.
Bibtex Entry:
@TechReport{TR100,
  author =	 "Florian Seitner",
  title =	 "Robust detection and tracking of objects",
  institution =	 "PRIP, TU Wien",
  number =	 "PRIP-TR-100",
  year =	 "2005",
  url =		 "ftp://ftp.prip.tuwien.ac.at/pub/publications/trs/tr100.pdf",
  abstract =	 "Due to the increasing availability of fast and cheap
                  hardware in the past few years, today a wide range
                  of complex visual tracking tasks is
                  possible. Efficient mathematical methods can provide
                  a high robustness which also makes visual tracking
                  interesting for many industrial purposes. However,
                  the high demands on quality and speed still provide
                  a major challenge for each tracking application. In
                  this thesis a tracking system is introduced, which
                  tries to address both demands appropriately by using
                  currently available algorithms to quickly track
                  pedestrians in video streams. By combining these
                  well-proved algorithms, a good solution regarding
                  computational complexity, accuracy and stability is
                  obtained. To achieve this task, a fast object
                  detector similar to the approach of Viola et
                  al. [Viola 2003] is used as one component in this
                  tracking system. This detector uses Haar-like
                  features which are very fast to compute and makes a
                  quick pedestrian detection in a frame possible. Next
                  to the detection system, an adaptive background
                  model sub-divides each frame into foreground and
                  background regions. As a compromise between
                  complexity and robustness a single-mode parametric
                  background model based on normal distributions and
                  wrapped normal distributions is used. Both
                  background model and detector are combined to
                  provide the tracking system with locations of
                  pedestrian-like regions and to sub-divide the body
                  into three parts: head, upper body and lower
                  body. After this segmentation into finer tracking
                  units a set of colour and spatial features for
                  further tracking is extracted from each part
                  individually. Individual and spatially separated
                  body parts also provide the possibility to use
                  colour histograms in a spatial sense. Moreover, an
                  appearance model provides accurate solutions and
                  approximations when occlusions or missing detections
                  occur.",
}
Powered by bibtexbrowser