Abstract: The problem of an object which is similar to the background texture or the boundary which is unclear in the image segmentation cannot be solved easily by the traditional video matting algorithm. In order to solve this problem, this paper proposes a real-time automatic algorithm of video matting based on information fusion of visual sensor and laser radar. The algorithm obtains the deep information from the original LIDAR point cloud data, which is used as prior knowledge to the improvement of spectral matting. It can simultaneously create a region of interest depth matting Laplacian matrix（DML）. Finally, the matting is gotten by cluster optimal iterative algorithm, and steerable filters is used to post-processing. Compared with the traditional algorithm of information fusion depth and no other information fusion algorithm, experiments in this paper show that the algorithm reduces the over-segmentation rate and improves the efficiency, and that the edge information of the matting object is more plump and smooth.

Key words: complex scenes, point cloud processing, depth information, spectral matting

摘要： 针对复杂场景下传统的视频抠像算法对目标物体与背景纹理相似或边界不清晰的图像分割困难的问题，提出了一种基于视觉传感器和激光雷达信息相融合的视频实时抠像算法。该算法从原始激光雷达点云数据中获取感兴趣区域深度信息，并作为先验知识融合到改进的谱抠图算法，创建感兴趣区域深度抠图拉普拉斯矩阵，通过聚类算法最优迭代得出抠像结果，并运用导向滤波器对抠像结果进行后处理。实验证明，对比于融合深度信息的传统算法和没有融合其他信息的算法，该算法降低了欠分割率、提高了运行效率，抠像目标的边缘信息也更加饱满、清晰、平滑。

关键词: 复杂场景, 点云处理, 深度信息, 谱抠图