基于宏块特征量化的视觉自适应实时监控方法

Macroblock feature quantification based real-time adaptive visual surveillance

提出一种新的基于宏块特征量化的视觉自适应实时监控方法,它采用考虑宏块特征空域关系和类似竞争分类的消除干扰策略,由宏块特征量化将对像素点的操作转变为对宏块直方图特征点的操作,通过兴趣点的提取使得对每帧的计算速度成数量级地提高.研究表明,该方法仅使用高、中、低3种外部阈值设置,便可使监控模型在几乎无参数调节下稳定运行,在动态图象序列中,能有效消除日光灯频闪光线、部分运动阴影以及局部小扰动带来的区域干扰.实验结果显示,该方法能够实现对安全防范中闯入类的有效监控,且目标跟踪快捷稳定,具有较高的工程实用价值.

A real-time visual surveillance method is presented, which is based on macroblock features quantification. The frame is divided into tiles and their macroblock features are extracted. The pixels operation is transferred to a nested operation of their macroblock feature points. The differences of feature frames are checked to find out residual motion region. The research shows that our novel method is quite effective and efficient in eliminating the region noise caused by fluorescent lamp, local moving shade and dithering in dynamic image sequences. Critical system parameters, such as threshold and processing regions, are adaptive to vary frame rates and computational constraints. According to different environment, it can track motion stably in real-time through a simple choice of high, median or low threshold setting. The experimental results show the promising performance in real-time visual surveillance in a changing environment.