结合表观、运动和边缘结构信息的遮挡边界检测方法  

Occlusion Boundary Detection by Combining Appearance,Motion and Edge Structure Cues in Video

作　　者：张世辉[1,2] 张红桥[1] 韩德伟[1] 

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]河北省计算机虚拟技术与系统集成重点实验室,河北秦皇岛066004

出　　处：《计量学报》2016年第3期241-245,共5页Acta Metrologica Sinica

基　　金：国家自然科学基金(61379065);河北省自然科学基金(F2014203119)

摘　　要：为了准确检测视频序列中的遮挡边界，提出了一种利用随机森林分类器检测遮挡边界的方法。该方法首先分割视频序列某一帧的边缘检测结果得到超像素和超像素边缘，并将超像素边缘分解为多个短的直线段。然后结合表观、运动和边缘结构信息提取每个直线段的遮挡相关特征并组合成特征向量，将此特征向量输入遮挡边界分类器检测每个直线段是否为遮挡边界。最后可视化所有确定为遮挡边界的直线段得到该帧图像的遮挡边界检测结果。实验结果表明所提方法具有较高的准确性。To detect the occlusion boundary in video sequences accurately, an occlusion boundary detection approach based on random forests classifier is proposed. Firstly, the edges of current frame in a video are segmented to obtain superpixels and superpixel edges, and then the superpixels edges are decomposed into short line fragments. Secondly, the occlusion related features of each line fragment are extracted by combining appearance, motion and edge structure cues and the extracted features are assembled to feature vector. After that, the feature vector of each line fragment is inputted to the occlusion boundary classifier to detect whether each line fragment is an occlusion boundary or not. Finally, the occlusion boundary of the current frame in a video is obtained by visualizing all the line fragments which belong to occlusion boundary. The experimental results show that the proposed approach has higher accuracy.

关 键 词：计量学 遮挡边界 随机森林 视频序列 超像素 直线段 光流差异 边缘结构 

分 类 号：TB96[机械工程—光学工程]