Citation Context

...l. [59] incorporate an additional hidden process for occlusion into a dynamic Bayesian network and rely on the statistical inference of the hidden process to reveal occlusion relations. Senior et al. =-=[52]-=- use appearance models to localize objects and use disputed pixels to resolve their depth ordering during occlusions. However, the system cannot maintain object identity after occlusions. Jojic and Fr...

Citation Context

...tly applicable to video object segmentation and tracking. Recently, a number of video segmentation measures have been proposed based on ground-truth (GT) segmentation maps. For example, Senior et al. =-=[8]-=- employed trajectories of the centroid of tracked objects and their velocities. Erdem et al. [9] utilize a combination of misclassification penalty, shape penalty, and motion penalty to assess the vid...

Citation Context

...he similarity of the trajectories. Any additional analysis, such as action recognition, event detection, etc., highly depends on the accuracy of the similarity assessment. Most existing measures [2], =-=[6]-=- compute a mean distance of the corresponding positions of two equal duration trajectories. Supplementary statistics such as variance, median, minimum, and maximum distances are also suggested to exte...

Citation Context

... we only summarize some of the basic principles of the image analysis. Readers are referred to papers about the PeopleVision system for more details of our implementation of video analysis algorithms =-=[18, 35, 36]-=-. Detection of objects of interest can be accomplished by one of two strategies. In a generic object detection approach, all objects of interest are dened in terms of one or more attributes (features...

Citation Context

...ch as head, torso, hands, etc. [18]. Research on segmentation or tracking of multiple people has focused on the analysis of the whole body in terms of the silhouettes [19,14], contours [20,21], color =-=[22]-=-, or blob [13,23]. The objective of this paper is to present a method for segmentation and tracking of multiple body parts in a bottom-up fashion. The method is a bottom-up approach in the sense that ...

Citation Context

...l. [18] incorporate an extra hidden process for occlusion into a dynamic Bayesian network, and rely on the statistical inference of the hidden process to reveal the occlusion relations. Senior et al. =-=[13]-=- use appearance models to localize objects and use disputed pixels to resolve their depth ordering during occlusions. However the system cannot maintain object identity after occlusions. Koller et al....

Citation Context

...e, to classify any pixel in the vicinity of the occlusion region as belonging to exactly one of the occluding objects. Most systems rely on the appearance features of objects to classify these pixels =-=[12, 43, 25, 19, 40]-=-. One particularly useful feature is the relative depth between occluding objects. Relative depth (or depth for short) can be determined in various ways. Elgammal and Davis [12] determine the depth of...

Citation Context

...luding objects (people) as a connected blob. They first recognized thesnumber of people in the blob using the number of heads, which come from the number ofslocal maxima in the contour. Senior et al. =-=[40]-=- tracked moving humans, bicycles andsvehicles using perspective images. They first took background subtraction to extract objects,sthen made temporal templates similar to Haritaoglu’s work, and divide...

Citation Context

...ion of the target in the occluded view can be inferred with the homography and the tracking result of the un-occluded view. Some work has been done in handling occlusion for both single view tracking =-=[9, 10] and-=- multi view tracking [1, 2, 3]. In [9], an appearance model is used to accomplish tracking. When occlusion is detected, the ”disputed” pixels are classified using a maximum likelihood classifier t...