Citation Context

...inputs. For multi-class recognition problems in computer vision, posterior probabilities are widely used to make decisions in applications such as pedestrian classification and orientation estimation =-=[18]-=-, image retrieval [15], attribute fusion [42], part-based human tracking [47], large-scale multiclass object categorization [4], and activity recognition [39], among others. To operate in open set sce...

Citation Context

...ors [18], [19]. Targeting its use on a real developmental robotic platform, in this article we apply it to a challenging visual discrimination task of namely real-world pedestrian pose classification =-=[20]-=-, [21] (see figure 1). Moreover, we illustrate the importance of the modulation mechanism by proposing various predictability measures and using simple bio-inspired projection and prediction algorithm...

Citation Context

... 218197, the ADABTS project. II. PREVIOUS WORK Estimating person orientation has been addressed in multiple ways. Several 2D single frame approaches combine orientation-specific person detectors [1], =-=[2]-=-, [3]. Work by [4] estimates body and head pose in batch mode, coupling the output of underlying classifiers. In [5], a texture sampled from a cylinder surrounding the person is shifted along the rota...

Citation Context

...of the camera: estimating the visual focus of attention for marketing strategies and effective advertisement methods (Ozturk et al., 2009), clothesshopping (Zhang et al., 2008), intelligent vehicles (=-=Enzweiler and Gavrila, 2010-=-), perceptual interfaces, etc. 2.2 Related work The different existing methods that estimate the orientation differ in several aspects: number of cameras and viewpoints, nature of the input (image, or...

Citation Context

...us on techniques for pedestrian state estimation and path prediction. For vision-based pedestrian detection, see recent surveys e.g. [10,12]. For pedestrian head/body orientation estimation, see e.g. =-=[5,13,14]-=-. State estimation in dynamical systems often involves the assumption that the underlying model is linear and that the noise is Gaussian, mainly due to the availability of the 620 J.F.P. Kooij et al. ...

Citation Context

...aper). II. RELATED WORK An extensive amount of research, meanwhile, exists on articulated pose estimation. Here, we focus on previous work on head and body orientation estimation. Approaches in body part orientation estimation are largely application dependent (see surveys [6] and [7] for head orientation estimation). Applications in human–machine interaction [8], [9] or entertainment [10] typically consider high-resolution images and cooperative subjects under controlled backgrounds to estimate the body part orientation. Applications in surveillance [11]–[16] and in intelligent vehicle [17], [18] domains need to cope with low-resolution images, with complex and dynamic backgrounds, and changing lighting conditions. To cope with these challenges, approaches often use robust lower level image features such as scale invariant feature transform and/or HOG [11], [12], [18]–[20], Haar [17], [21], local receptive fields (LRFs) [18], and distance metrics [9], [13], [19] in combination with different classification schemes (e.g., support vector machines (SVMs) [12], [13], [18], [20], [21], neural networks (NNs) [18], random regression/decision trees or ferns [9], [11], [19], or boosting cascad...

Citation Context

...utomatically yield an assumption of the orientation of a pedestrian in discrete steps. Enzweiler et al. use a similar approach but try to estimate a continuous orientation from the classifier outputs =-=[15]-=-. For this, they again use discrete orientation steps to train their view dependent detectors. The outputs of these classifiers are used as weights for a Gaussian mixture. Zou et al. propose a color b...