We propose an approach to the segmentation of video objects based on motion cues. Motion analysis is performed by estimating local orientations in the spatio-temporal domain using the three-dimensional structure tensor.
These estimates are integrated as an external force into an active contour model, thus stopping the evolving curve when it reaches the moving objects' boundaries. To enable simultaneous detection of several objects we reformulate the tensor-based active contour model using the level-set technique. In addition, a contour refinement technique has been developed to approximate the real boundaries of the moving objects.
We have developed a second approach for object segmentation in historical videos (see second demo). This approach identifies point-correspondences in shots and applies a robust parameter estimation technique based on the RANSAC algorithm to identify camera parameters. A background image based on the median of the transformed video frames is constructed and each frame is compared with this background image.
Segmentation of objects in historical videos (for black & white videos; no tracking or validation step were applied in this demo): Short Demo Video (historical video: 22MB) (segmented objects video: 4MB) (demo program for MS windows: 4MB)