基于超像素时空特征的视频显著性检测方法

Video Saliency Detection Method Based on Spatiotemporal Features of Superpixels

提出一种基于超像素时空特征的视频显著性检测方法。所提方法可对图像进行超像素分割,提取颜色梯度和运动梯度特征,以构建超像素级时空梯度图。用平均加权测地距离来衡量时空梯度图上每一个超像素相对于其邻域的时空显著程度,形成时空显著图。根据时间域上目标运动的连续性,并借助熵的概念来表征运动模式的一致程度,构建运动一致性图。融合时空显著图和运动一致性图,通过自适应阈值处理定位运动目标。实验从可视化分析和定量评估两个方面将所提方法与其他算法进行对比,结果表明所提方法具有较强的抗环境干扰能力,适用于背景纹理复杂或环境随机变化的视频中运动目标的检测,其检测准确率高达92%。

A novel video saliency detection method is proposed based on the spatiotemporal features of superpixels,which is used to the superpixel segmentation of images and extract the features of color gradient and motion gradient for the construction of a spatial-temporal gradient map of superpixels. The average weighted geodesic distance is used to measure the spatiotemporal saliency degree of each superpixel relative to its neighbor on the spatiotemporal gradient map, and thus the spatiotemporal saliency map is formed. In order to obtain the motion coherency map,the motion entropy in the multiple continuous frames is computed to represent the motion coherence of motion object over time. The fusion of spatiotemporal saliency maps and motion coherency maps is applied to locate in the salient motion using adaptive segmentation. In addition, the performance of the proposed method is compared with those of the other algorithms in experiments from two perspectives of visual analysis and qualitative evaluation. The results show that the proposed method is robust and suitable for the detection of moving targets in videos within complex background texture and changeable environment. Moreover, the detection precision is up to 92%.