基于时空模型的快速视频显著区域检测

Fast Visual Attention Region Detection in Video Using Spatiotemporal Model

传统的视频显著区域检测方法运算量大且难以处理包含复杂运动的视频,本文提出一种基于时空模型的快速显著区域检测方法。该方法用一种新的模糊聚类算法对特征点的运动轨迹进行无监督聚类,对不同运动类型进行分类。在复杂运动情况下,该算法通过计算优化的聚类中心的个数以获得运动类型数,再将异常数据剔除后,生成运动显著图。而在空间显著区检测方面,则利用反差模型以及Gabor滤波器获得图像的静态显著图。在此基础上,还提出一种基于运动优先思想的时空混合方法,将运动和空间显著图动态合成视觉显著图。实验证明,该方法能够有效地提取视频显著区域,与传统的方法相比该方法平均耗时更少且更方便。

A novel fast visual attention map detection approach using spatiotemporal model is proposed. To generate motion saliency map, the feature point motion vectors are extracted as motion feature. Then a new fuzzy cluster method is used for cluster validation to analyze motion consistency. The spatial saliency map is generated using Gabor filter and center-surround descriptor. Finally, the motion and the spatial saliency map in motion fashion are synthesized to creat the overall spatiotemporal attention map. Experimental results show that the method can achieve both good accuracy and the real-time performance.