February 2012
Spotlight Summary by Kedar Khare
Tracking of multiple objects in unknown background using Bayesian estimation in 3D space
Automated tracking of objects in an image scenery is an important practical problem for remote sensing, surveillance, security imaging, robotic vision and related areas. While it may seem straightforward to track an isolated object in a sequence of images, the problem becomes challenging when there is a complex variable background and the objects of interest are occluded as they move. In general it is clear that segmentation and tracking performance will improve if we have a multi-dimensional representation for the objects and their properties.
In the article by Zhao, Xiao, Cho and Javidi, the authors report the performance of their scheme for tracking multiple objects in a 3D scene. As it may be clear, 3D imaging information already provides an additional dimension of information compared to 2D imaging. The authors achieve 3D imaging by the well-known method of integral imaging first studied by Lippmann over 100 years ago. In this modality, multiple images of the scene are simultaneously recorded by an array of imaging devices, such that each image provides a slightly different perspective. In a computational version of this method, the recorded images are back-projected to get the appropriate objects in focus in each plane in the depth direction, the remaining background resulting in a blur. The authors use statistical modelling of background and objects in each plane in the depth direction to segment the objects that are then tracked in the video scenery. In particular, they model the background with a Gaussian distribution and the objects of interest with a suitably chosen Gamma distribution. Local image statistics then allows them to distinguish between the objects of interest and the background. Superior performance of this method as compared to tracking of objects in the corresponding 2D scenery is illustrated in this paper. The results of this technique appear to be independent of location, rotation, scale of the object as well as illumination conditions, and can potentially lead to many interesting applications.
You must log in to add comments.
In the article by Zhao, Xiao, Cho and Javidi, the authors report the performance of their scheme for tracking multiple objects in a 3D scene. As it may be clear, 3D imaging information already provides an additional dimension of information compared to 2D imaging. The authors achieve 3D imaging by the well-known method of integral imaging first studied by Lippmann over 100 years ago. In this modality, multiple images of the scene are simultaneously recorded by an array of imaging devices, such that each image provides a slightly different perspective. In a computational version of this method, the recorded images are back-projected to get the appropriate objects in focus in each plane in the depth direction, the remaining background resulting in a blur. The authors use statistical modelling of background and objects in each plane in the depth direction to segment the objects that are then tracked in the video scenery. In particular, they model the background with a Gaussian distribution and the objects of interest with a suitably chosen Gamma distribution. Local image statistics then allows them to distinguish between the objects of interest and the background. Superior performance of this method as compared to tracking of objects in the corresponding 2D scenery is illustrated in this paper. The results of this technique appear to be independent of location, rotation, scale of the object as well as illumination conditions, and can potentially lead to many interesting applications.
Add Comment
You must log in to add comments.
Article Information
Tracking of multiple objects in unknown background using Bayesian estimation in 3D space
Yige Zhao, Xiao Xiao, Myungjin Cho, and Bahram Javidi
J. Opt. Soc. Am. A 28(9) 1935-1940 (2011) View: Abstract | HTML | PDF