確率的モデリングに基づくRGB画像列の運動領域の分割

概要

論文の詳細を見る
One of the possibilities of motion segmentation is labeling approach, i.e. assigning to each pixel one of two labels : motion, or background. Motion segmentation in RGB sequences has to be based on an extended local information vector comparing with the analogous algorithms for monochromatic images. In our method the local information vector : o_k(i)=[ΔR_k(i), ΔG_k(i), ΔB_k(i), R_k(i), G_k(i), B_k(i)]^T(1) is used where, R, G, B mean color components, ΔR_k(i)=R_(k+1)(i)-R_k(i), k corresponds to the frame number, and i points a pixel in a frame. Let us consider two consecutive sequence frames. In such a case the motion of any object results in four different areas of changes taking place in both images : covered background, recovered background, static background, and overlapping area. In fact, the last type exists if the object motion is not too fast. If there is not the overlapping situation, the analysis of three consecutive sequence frames is enough to perform the motion segmentation in the middle one, however, One cannot assume such a case usually.
社団法人電子情報通信学会の論文
1996-03-11