针对伪装问题的多信息融合改进ViBe算法

Improved visual background extractor algorithm using top-down information for camouflage challenge

视觉背景提取器(ViBe)算法是一种经典的运动目标检测算法,但ViBe算法在面对伪装问题时性能大幅下降。为了更好地解决伪装问题,文中提出了一种基于权重隶属度函数的融合自顶向下信息和自底向上信息以增强前景检测的改进ViBe算法。首先,背景模型用于捕获自底向上信息,而前景模型用于捕获自顶向下信息。然后,根据像素与每个模型的匹配样本的数量和像素与其匹配样本之间的距离分别计算像素与上述两个模型的匹配度。最后,使用基于权重隶属度的同时使用前景模型匹配度和背景模型匹配度的决策框架来确定像素标签。实验数据表明,提出的方法在主观评价和客观评价方面均优于文中采用的其他对比算法,并且,提出的算法相比于原始ViBe算法在主观评价和客观评价上均有明显提升。实验结果证明了提出的方法的有效性,特别是处理存在伪装问题的序列。

The visual background extractor(ViBe)algorithm is a classic moving object detection algorithm.However,the ViBe algorithm shows strong decrease on performance when faced with a camouflage challenge.To better handle camouflage problem,both bottom-up cues and top-down cues are used in ViBe algorithm to augment foreground detection through weighted membership function.Firstlt,background model is used to capture bottom-up cues,while foreground model is used to capture top-down cues.Then,the pixel matching degree with two above models is calculated,respectively according to the number of matching samples with each model and distance between the pixel and its matching samples.Finally,a membership-based decision framework using foreground model matching degree and background model matching degree is used to determine the pixel label.Experimental data show that the algorithm is superior to other comparison algorithms in the paper in the subjective evaluation and the objective evaluation,and the algorithm is significantly improved in the subjective evaluation and the objective evaluation compared with the original ambience algorithm.Experimental results demonstrate the effectiveness of the algorithm,especially for sequences faced with camouflage challenge.