S. McKenna and S. Gong. Tracking Faces. In Proceedings of the Second International Conference on Automatic Face and Gesture Recognition FG'96, pages 271--276, Killington (VT), USA, October 1996. IEEE.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....feature points on different faces are required and as a result, real time performance is obtained. The models are constructed from aligned image data labelled with pose angles. Efficient focus of attention based on colour and motion cues is used to bootstrap face image search in the image plane [6, 7]. 2 Acquisition of Labelled Views across the View sphere In order to build appearance models, example views labelled with 3D pose angles (both tilt and yaw) are required. A system was designed that utilises both a magnetic sensor attached to the subject s head and a camera calibrated relative to ....

....training data. In practice, the number of examples available at each view is small. Alternatively, appearance models based on similarity vectors to a limited number (in tens) of prototype faces at multiple views can be adopted. Given that face images at the frontal view can be readily detected [6], let a similarity vector # to prototypes for a detected face image at the frontal view be measured using Equation (1) Pose recognition and tracking can then be performed by finding the next pose # (both yaw and tilt) which maximises ######## # # ## ##### # # # # ### # (4) where ### # ....

S. McKenna and S. Gong, "Tracking faces," in IEEE Int. Conf. on Automatic Face and Gesture Recognition, Killington, Vermont, U.S., October 1996, pp. 271--277.

....for matching. A second motivation to further analyze the shape is that adaptive background separation in complex scenes detects recently displaced objects. Too identify parts of the human body feature detectors that find these parts directly can be constructed. For finding faces neural nets [10, 11] have been used. A different approach uses edges in gradient images [12] For high resolution face images recognition can be further enhanced by using active contour models [13] Templates or sets of templates can equally be used for finding faces however robust person independent location under ....

S. McKenna and S. Gong, "Tracking faces," in International Conference on Automatic Face and Gesture Recognition, pp. 271--276, Oct. 1996.

....Three people trying to steal the ellipse from the subject Figure 6: Demonstration of the tracker s performance in various situations. These and other MPEG sequences are available from http: vision.stanford.edu birch. the background, all simultaneously. Template and neural network based trackers [6, 9, 11, 18], as well as trackers based on facial color [4, 5, 9, 15, 16, 18] cannot handle severe out of plane rotation because such a rotation causes the face to disappear. The colorbased techniques also tend to have difficulty with skincolored objects or other people in the background. Trackers utilizing ....

....color [4, 5, 9, 15, 16, 18] cannot handle severe out of plane rotation because such a rotation causes the face to disappear. The colorbased techniques also tend to have difficulty with skincolored objects or other people in the background. Trackers utilizing some form of background differencing [5, 10, 11, 12, 18, 19, 20] either require a static camera or restrict the camera s motion to rotation about its focal point. 3 Moreover, many of these techniques perform motion based figure ground segmentation, which tends to fail when the camera zooms or when multiple objects move 3 In [10] the camera may move ....

S. McKenna and S. Gong. Tracking faces. In Proc. of the Second International Conference on Automatic Face and Gesture Recognition, pages 271--276, 1996.

....is used to perform the experiments and demonstrate that our approach is feasible and promising. 1. Introduction Due to its potential for surveillance, security and human computer interface, tracking human motion with computer vision techniques has become a popular research field [1] 4] 6] 9] 10][12]. In this field, many researchers are specifically interested in tracking human heads or faces [1] 6] 9] 12] For the problem of human head tracking, two important issues should be addressed: what to track (i.e. detection of the human head) and how to track. Because it is difficult to ....

.... Due to its potential for surveillance, security and human computer interface, tracking human motion with computer vision techniques has become a popular research field [1] 4] 6] 9] 10] 12] In this field, many researchers are specifically interested in tracking human heads or faces [1] 6] 9][12]. For the problem of human head tracking, two important issues should be addressed: what to track (i.e. detection of the human head) and how to track. Because it is difficult to automatically detect human heads or faces in images having complex backgrounds, much previous research either bypassed ....

S. McKenna and S. Gong, "Tracking Faces", 2nd International Conf. on Automatic Face- and Gesture- Recognition, Killington, Vermont, 1996, pp. 271-276.

....commonly used for dealing with feature variations: correlation templates [8, 9] deformable templates [10] spatial image invariants [11] and neural networks [2, 3] These methods are, however, computational expensive and hardly achieve real time performance. For example, the system described in [12] tracks object at about 5 frames second speed with a 189 x 144 image by using a neural network to detect faces. Color is another feature on human faces. Using skin color as a feature for tracking a face has several advantages. Processing color is much faster than processing other facial features. ....

S. McKenna and S. Gong, "Tracking faces," Proc. the 2nd Int. Conf. on Automatic Face and Gesture Recognition, Killington, Vermont, 1996.

....feature points on different faces are required and as a result, real time performance is obtained. The models are constructed from aligned image data labelled with pose angles. Efficient focus of attention based on colour and motion cues is used to bootstrap face image search in the image plane [6, 7]. 2 Acquisition of Labelled Views across the View sphere In order to build appearance models, example views labelled with 3D pose angles (both tilt and yaw) are required. A system was designed that utilises both a magnetic sensor attached to the subject s head and a camera calibrated relative to ....

....training data. In practice, the number of examples available at each view is small. Alternatively, appearance models based on similarity vectors to a limited number (in tens) of prototype faces at multiple views can be adopted. Given that face images at the frontal view can be readily detected [6], let a similarity vector to prototypes for a detected face image at the frontal view be measured using Equation (1) Pose recognition and tracking can then be performed by finding the next pose (both yaw and tilt) which maximises .E] 4 XDX XDX acb , 1 K d (4) XPX ....

S. McKenna and S. Gong, "Tracking faces," in IEEE Int. Conf. on Automatic Face and Gesture Recognition, Killington, Vermont, U.S., October 1996, pp. 271--277.

....estimating their aspect, two factors are exploited. First, simple attention focusing cues are used to localise potential facial regions. Efficient focus of attention based upon motion and colour cues has been used to direct face search with a generic appearance based neural network face detector [12,18]. Second, temporal cor relation of head pose and face position are exploited by tracking the face and its pose, rather than searching for these parameters directly in each video frame. To associate moving faces in real time, we adopt a view based approach that utilises learnable appearance models ....

....training data. In practice, the number of examples available at each view is small. Alternatively, appearance models based on similarity vectors to a limited number (in tens) of prototype faces at multiple views can be adopted. Given that face images at the frontal view can be readily detected [12], let a similarity vector ct to prototypes for a detected face image at the frontal view be measured using Equation (2) Pose recognition and tracking can then be performed by finding the next pose 0 (both yaw and tilt) which maximises c(0) 11411 h(4, t x) 6) where IIctll is the L2 norm of ....

S. McKenna and S. Gong. Tracking faces. In IEEE International Conference on Face Gesture Recognition, pages 271-277, Killington, Vermont, U.S., October 1996.

....bootstrap the tracking and pose estimation process. Methods for face detection usually assume frontal or near frontal views and tend to be computationally expensive. They have been based upon colour [13] silhouette, spatial configuration of facial features and pattern recognition techniques (see [14] for references) Face detection is not the main topic of this paper and the systems described here used an appearance based matching scheme for detection and tracking. Further references on methods for face processing can be found in reviews [15] 16] DRAFT May 13, 1998 5 A. Feature based ....

....used to investigate the role of GWP face representations. This in turn motivates the design of the real time systems for pose estimation. Sequences of heads rotating from profile to profile under different lighting conditions were obtained as output from a head tracking system described elsewhere [14], 29] These were 60 frames long and were automatically normalised with respect to translation and scale by the tracker. Each image was also normalised by subtracting its mean intensity and dividing by its standard deviation. This corrected variations in overall illumination intensity, camera ....

S. J. McKenna and S. Gong, "Tracking faces," in Proc. 2nd Int. Conf. on Automatic Face and Gesture Recognition, Killington, VT., 1996.

....can be seen in Figure 2. In essence, this process is a closed loop module that includes the computation and fusion of three different visual cues: motion, colour and face appearance models. Face tracking based upon motion and a face appearance model has been addressed in greater detail elsewhere [1,2]. The use of colour is described here. The remainder of this paper then focuses upon person identification within such a framework. Complementary to recognition, appearance based mechanisms for real time face pose estimation have been developed which can be used to improve the robustness of ....

....space provides appearance based models of identity suited to all four tasks. Gaussian mixtures are then presented and evaluated for this purpose. Conclusions are drawn in section 6. 2 Locating and tracking faces using colour A system for detecting and tracking faces was previously described [1,2]. It combined motion detection by spatio temporal filtering with an appearancebased face model in the form of a neural net. Multiple person tracking was performed using time symmetric matching and Kalman filtering. In this section, the use of colour as a cue for detection and tracking is ....

S. J. McKenna and S. Gong, "Tracking faces," in Proc. 2nd Int. Conf. on Automatic Face and Gesture Recognition, Killington, Vermont, US, October 1996.

....the face. The remainder of this paper describes the system in more detail. Some examples are given to illustrate the performance of the tracker. 2 Tracking Faces This section describes the system for tracking peoples faces. More detailed descriptions of some aspects can be found elsewhere [5, 6]. 2.1 Tracking multiple motions Visual motion is estimated by convolving the intensity history of each pixel I(x; y; t) with the second order temporal derivative of a Gaussian function G(t) yielding an image of temporal zero crossings S(x; y; t) 2] S(x; y; t) 2 G(t) t 2 I(x; y; ....

....a crude localisation of the human head but an accurate and robust face tracking system requires further knowledge which is provided here in the form of a face appearance model implemented using a neural network. Rowley et al. describe a neural network for face detection in static scenes [7] See [5] for a discussion and references on face detection) A similar approach is used here to localise and track faces in dynamic scenes. Localised connectivity (receptive elds) can be used to constrain the network, yielding improved generalisation. Radial basis function networks (RBFN) were also used ....

S. McKenna and S. Gong. Tracking faces. In IEEE Second Internatioanl Conference on Automatic Face and Gesture Recognition, Killington, Vermont, US, October 1996.

....feature points on different faces are required and as a result, real time performance is obtained. The models are constructed from aligned image data labelled with pose angles. Efficient focus of attention based on colour and motion cues is used to bootstrap face image search in the image plane [6, 7]. 2 Acquisition of Labelled Views across the View sphere In order to build appearance models, example views labelled with 3D pose angles (both tilt and yaw) are required. A system was designed that utilises both a magnetic sensor attached to the subject s head and a camera calibrated relative to ....

....training data. In practice, the number of examples available at each view is small. Alternatively, appearance models based on similarity vectors to a limited number (in tens) of prototype faces at multiple views can be adopted. Given that face images at the frontal view can be readily detected [6], let a similarity vector ff to prototypes for a detected face image at the frontal view be measured using Equation (1) Pose recognition and tracking can then be performed by finding the next pose (both yaw and tilt) which maximises L( jjff t jj h(ff t ; ff t Gamma1 ) 4) ....

S. McKenna and S. Gong, "Tracking faces," in IEEE International Conference on Automatic Face and Gesture Recognition, Killington, Vermont, U.S., October 1996, IEEE, pp. 271--277.

....background sequence was performed. Only pixels inside the search area of the tracker were classified. All pixels outside this area were rendered as background. 7 Focus of Attention for Face and Gesture Recognition A system for tracking multiple objects based on motion was reported at FG 96 [5]. A similar approach is adopted here for detecting and tracking multiple objects based on their colour. In particular, a skin colour model can be used to efficiently focus attention on faces and hands for face and gesture recognition processes. Figure 5 shows such a system running on a 200MHz PC ....

S. McKenna and S. Gong. Tracking faces. In 2nd Int. Conf. Face and Gesture Recognition, pages 271--276, 1996.

No context found.

S. McKenna and S. Gong. Tracking Faces. In Proceedings of the Second International Conference on Automatic Face and Gesture Recognition FG'96, pages 271--276, Killington (VT), USA, October 1996. IEEE.

No context found.

S. McKenna and S. Gong, "Tracking faces", Proc. 2 ICAFGR, Killington, Vermont, US, 1996. 47

No context found.

McKenna, S. J. and Gong, S. Tracking faces. in Proceedings of the Second International Conference on Automatic Face and Gesture Recognition. IEEE Comput. Soc. Press.

No context found.

S. McKenna and S. Gong, "Tracking faces", Proc. 2 ICAFGR, Killington, Vermont, US, 1996. 47

No context found.

S. McKenna and S. Gong, "Tracking Faces," Proc. Second Int'l Conf. Automatic Face and Gesture Recognition, pp. 271-276, 1996.

No context found.

McKenna, S. J. and Gong, S. Tracking faces. in Proceedings of the Second International Conference on Automatic Face and Gesture Recognition. IEEE Cornput. Soc. Press.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC