Abstract
In this paper, we have addressed mainly a detection based method for video events detection. Harris points of interest are tracked by optical flow techniques. Tracked interest points are grouped into several clusters by dint of a clustering algorithm. Geometric means of locations and circular means of directions as well as displacements of the feature points of each cluster are estimated to use them as the principle detecting components of each cluster rather than the individual feature points. Based on these components each cluster is defined either lower-bound or horizon-bound or upper-bound clusters. Lower-bound and upper-bound clusters have been used to detect potential ordinary video events. To show the interest of the proposed framework, the detection results of ObjectDrop, ObjectPut, and ObjectGet atTRECVid 2008 in real videos have been demonstrated. Several results substantiate its effectiveness, while the residuals gives the degree of the difficulty of the problem at hand.