147. A DISCRIMINANT CENTER-SURROUND MOTION SALIENCY ALGORITHM

Department: Electrical & Computer Engineering
Faculty Advisor(s): Nuno Vasconcelos

Primary Student
Name: Vijay Mahadevan
Email: vmahadev@ucsd.edu
Phone: 858-534-4538
Grad Year: 2010

Abstract
A discriminant center-surround algorithm for identifying motion saliency in a video sequence is introduced as an extension to the already proposed biologically inspired discriminant saliency framework for static images. In general, discriminant saliency at a location is the power of a set of features to differentiate the region containing the location, i.e. the center, from its surround. By extending this principle to appearance and motion features, we obtain a measure of how distinct a moving object is from its surround. We choose two types of motion descriptors : optical flow and dynamic textures. The resulting motion saliency detector requires no training, and enables identification of salient objects in both spatial and temporal domains. Being a discriminant technique, the algorithm is invariant to the effect of egomotion.We demonstrate the application of this saliency detector to salient motion detection in the presence of egomotion and on challenging scenes with complex dynamic backgrounds. The algorithm is tested on real video clips and performance is evaluated using manually annotated groundtruth. It is shown that the proposed algorithm clearly outperforms the state-of-the-art techniques for such tasks. Applications for the proposed algorithm include motion compensation, background subtraction, and motion vector processing for scalable video coding .

« Back to Posters or Search Results