Change detection is an important process in many video-based applications such as anomaly event detection and video surveillance. This paper proposes a foveated vision framework that simulates the human visual system for change detection. It contains two phases—first identifying regions with visual changes due to significant motion, and then, the extraction of detailed information of the change. In phase I, change proposals (CPs) and background are segregated by analyzing the intensity and motion features. In phase II, visual changes are estimated from the CPs by analyzing the photometric and textural features. Each phase of analysis has a unique pre-generated archetype. A probabilistic refinement scheme is used to rectify the labeling of background and change. In each phase of analysis, the result is used to update the archetype immediately. Some well-known and recently proposed background modeling/subtraction algorithms are selected for our comparative study. Experimentations are performed on various video datasets. In some videos, our method can achieve higher accuracy than some recently proposed methods by 30%. In the large-scale experimentation using all the testing videos, our method can achieve higher average accuracy than the second best method by more than 3%.

中文翻译：

---

使用运动和纹理特征进行视觉变化检测的注视点视觉框架

变化检测是许多基于视频的应用中的一个重要过程，例如异常事件检测和视频监控。本文提出了一种模拟人类视觉系统进行变化检测的注视点视觉框架。它包含两个阶段——首先识别由于显着运动而产生视觉变化的区域，然后提取变化的详细信息。在第一阶段，通过分析强度和运动特征来分离变更建议（CP）和背景。在第二阶段，通过分析光度和纹理特征从 CP 估计视觉变化。分析的每个阶段都有一个独特的预先生成的原型。概率细化方案用于校正背景和变化的标记。在分析的每个阶段，结果用于立即更新原型。我们选择了一些著名的和最近提出的背景建模/减法算法进行比较研究。在各种视频数据集上进行了实验。在一些视频中，我们的方法可以达到比最近提出的一些方法高 30% 的准确率。在使用所有测试视频的大规模实验中，我们的方法可以达到比第二好的方法高 3% 以上的平均准确率。