复杂场景中基于对象的运动目标检测方法

Object-oriented Moving Target Detection in Complex Scenes

基于像素层面的混合高斯背景建模方法不能很好的解决动态背景中的运动目标检测问题。由于背景像素运动的复杂性,该方法很难将动态背景建入模型,会造成大量的误检。本文在混合高斯背景建模的基础上,通过空域和时域对动态背景产生的误检进行抑制。在空域运用MRF模型和混合高斯模型分别计算像素点的先验概率和类条件概率,通过结合像素点的先验概率和类条件概率完成前景图像的分割,在很大程度上去除了小面积的误检;在时域通过目标的运动持续性,运动显著性和面积变化稳定性三个目标特征过滤大面积的误检。通过实验表明,在保证较高检测精度的情况下,该方法能够在很大程度上抑制动态背景产生的误检。

Pixel-based MOG background modeling is not a good method to deal with moving target detection in dynamic background.It is hard to model the dynamic background in this method due to the complex movement of background pixels.Thus,huge sums of false positives would occur.This paper aims to suppress the false positives in space domain and time domain.In space domain,class-prior probability and class-conditional-probability of a pixel were calculated from MRF model and Mixture of Gaussians（MOG） model.By combining the class-prior probability and class-conditional-probability,foreground segmentation was completed to suppress a majority of false positives of small size.In time domain,three target characteristics were observed：motion constancy,motion saliency and area stabilization.The result of experiments show that the false positives on dynamic background could be suppressed to a great extent,meanwhile,good detecting accuracy is ensured.