融合运动信息的三维视觉显著性算法研究

Research on 3D Vision Saliency Algorithm with Fusion Motion Information

计算机视觉领域中的视觉显著性研究大多局限于二维图像层面,而忽略人的视觉注意力决策是在三维动态场景下发生的。为此,在融合多种特征的显著性计算框架基础上,提出一种三维视觉显著性算法。通过场景的颜色信息、运动信息和深度信息分别计算各个特征通道下的显著性结果,再经过动态的融合得到最终的显著性结果。同时针对三维场景下显著性数据集的稀缺问题,给出一个用于评价三维动态场景下显著性算法的数据集。与HC算法、RC算法、GMR算法的对比结果验证了该算法具有明显的优势,并且更符合人眼的视觉注意力机制。

Most of the research on visual saliency in computer vision is limited to 2D image,and ignore the fact that human vision attention decision making operates in real 3D dynamic scene. So this paper proposes a 3D vision saliency algorithm which combines saliency computing framework with multiple features. Using color information of the scene, motion information and depth information, the saliency results of each characteristic channel are calculated respectively, and the final saliency results are obtained by dynamic fusion. Due to the lack of datasets of saliency in 3D scenes,this paper also proposes a new dataset for the evaluation of relevant algorithms. Compared with HC algorithm,RC algorithm, and GMR algorithm,quantitative results prove that the proposed algorithm outperforms all the other algorithms and better fits the visual attention mechanism of human eyes.