Serdar
Ince

This work presents a simple approach for detection of occlusions based on a geometric principle.

A typical motion field computed under some form of spatial regularization will lead to converging motion vectors originating in occlusion areas of the reference frame (area A in Figure). Such vectors cannot provide a good intensity match and assume compromise coordinates with respect to the neighboring vectors from, e.g., a moving object and static background.

This convergent behavior is a compromise between the lack of intensity match and spatial smoothness enforced, and potentially leads to multiple vectors pointing to the same location in the target frame. This might suggest that a high spatial density of motion-compensated positions in the target frame (I2 in Figure) is indicative of an occlusion area. However, in practice, it turns out that results are very sensitive the selected density threshold. On the other hand, pixels in the target frame that did not exist in the reference frame (newly-exposed pixels in area B) have no relationship with the reference frame and, as such, cannot be pointed to by forward motion vectors.

Thus, areas in the target frame that are void of motion-compensated projections can be relatively easily detected. The exposure areas in I2, will be the occlusion areas when the motion vectors are pivoted on I2. This is the basis of the proposed detection algorithm. You can access our paper above for a detailed explanation.

S. Ince and J. Konrad , "Geometry-based estimation of occlusions from video frame pairs," in Proc. IEEE Int. Conf. Acoustics Speech Signal Processing, 2005, [abstract] [gzip-compressed PS: 941KB], [PDF: 752KB].

[Home] [Biography] [Research] [Publications] [Personal] [Downloads] [Turkce]