Full-motion recovery from multiple video cameras applied to face tracking and recognition

Abstract

Robust object tracking still remains a difficult problem in computer vision research and surveillance applications. One promising development in this area is the increased availability of surveillance cameras with overlapping views. Intuitively, these overlapping views may lead to more robust object tracking and recognition. However, combining the information from the multiple cameras in a meaningful way is challenging. Our contribution in this work is a novel approach to object tracking by robustly and accurately recovering the full motion of the object from multiple cameras. This is accomplished by explicitly fusing the information from multiple cameras into a joint 3D motion calculation. We apply this approach to the tracking of faces in multiple video cameras and utilize the 3D cylinder model to realize the motion calculation. The method is demonstrated on a sequence of real data for pose estimation of the face. Also, the 3D cylinder texture map from the tracking result is used in face recognition. The performance of fullmotion recovery from multiple cameras is shown to produce a significant increase in the accuracy of face pose estimation and results in a higher face recognition rate than from a single camera. Our approach may be applied to other types of object tracking such as vehicles.

Determination of the Three-Dimensional Rate of Cancer Cell Rotation in an Optically-Induced Electrokinetics Chip Using an Optical Flow Algorithm

“…Another common approach to 3D motion recovery is to choose an initial reference template and then match the images with this reference template. This method is widely used for recognition tracking and motion recovery of human facial and head gestures [ 20 , 21 ]. As the texture and the rotation axis are difficult to predict and there is no obviously-preferred direction for the cells, the template matching method is not feasible.…”

Dense-HOG-based drift-reduced 3D face tracking for infant pain monitoring

“…Others recover motion by Lucas-Kanade template alignment [9], e.g. with a cylinder [10] or ellipsoid head model and for multi-camera setups [11]. We adopt the same principle for our algorithm.…”

“…As a main innovation over [10,11] , we use dense-HOG feature vectors as appearance values instead of pixel intensities. We use 36-dimensional vectors, from a HOG variant with blocks of 2×2 cells, cells of 8×8 pixels, and 9-bin histograms for signed orientations.…”

“…In [10,11] , the principle of head pose recovery by 3D alignment is used with pixel intensity templates. This may give problems with less-pronounced texture and less-uniform illumination, as e.g.…”