Citation Context

...ractive alternative is to ignore dynamics and find people in each frame independently, using such cues as local motion [15] or appearance [11]. As far as we know, no current person tracker meets all our criteria. This is most likely because of difficulties with data association [2]; only a small part of the abundant image data contains any information about the person and one may need to determine which part this is. Particle filters do data association only implicitly which explains their attraction. One may use a variety of templates encoding the appearance of the person being tracked (e.g. [12, 16]). To date, these templates have been learned in advance of tracking, and so cannot use properties like the color of clothing, which change from instance to instance. People tend not to change “appearance” (color and texture of clothing, etc.) from frame to frame. We describe a people tracker which builds models of the people to be tracked from the video sequence and then tracks them. This has considerable advantages: First, knowing the appearance model of each body part greatly constrains our search and so simplifies data association. Second, we can prevent drift, recover from occlusion relat...

Citation Context

...th the motion. One approach to embed template matching in a probabilistic tracking framework was proposed for complete image frames by Jojic et al. [24] and for exemplar templates by Toyama and Blake =-=[43]-=-. However, it is acknowledged that “one problem with exemplar sets is that they can grow exponentially with object complexity. Tree structures appear to be an effective way to deal with this problem, ...

Citation Context

...DT-based matching; one need not match a template for each location, rotation, or other transformation. Exemplar-based shape representations have furthermore been applied to tracking. Toyama and Blake =-=[28]-=- devise a probabilistic framework, termed Metric Mixture, which they use for tracking human bodies and mouths. Stenger et al. [27] extend the hierarchical representation of [12] by means of a Bayesian...

Citation Context

...ed to combine visual cues such as edge and texture in a particle filter framework. Particle filters have furthermore been applied in combination with low-level [14,15], high-level [9], exemplar-based =-=[18]-=-, or mixed-level [11] object representations. In terms of representation, compact low-dimensional object parameterizations can be obtained by linear subspace techniques, e.g. using shape (PDMs) [3,9],...

Citation Context

...thus attempts to combine tracking methods that are based on dynamics-driven sample propagation and image likelihood measurements (e.g. [17,21]), with more independent detection style approaches as in =-=[23,18]-=-. c.f . [2]. An approach somewhat similar to ours, developed indenpendently and contemporaneously, is described in [19]. Organization: The next two sections discuss a mixture-ofGaussians based scheme ...

Citation Context

...lars have been used with success in a variety of vision tasks, including image synthesis [3, 4], super-resolution [5, 6], image and video completion [7, 8] and combinedy tracking and object detection =-=[9, 10]-=-. The use of exemplars is attractive for several reasons. A relatively small number of representative exemplars can capture high-order statistics, since each exemplar can simultaneously express depend...

Citation Context

...mination, ν = {gt, αt, βt}. We have so far neglected visibility effects (occlusions and cast shadows), but we will come back to it. Note that the equation above is reminiscent of deformable templates =-=[50, 9, 18]-=-, although here we do not know the templates. Correspondence is also naturally related to wide-baseline matching [43, 15, 15, 10, 28]. From now on we will restrict our attention to the model (2). 1.3 ...