基于像素自适应背景建模的运动目标分割

Moving object segmentation through pixel-based adaptive background modeling

针对PBAS(pixel-based adaptive segmenter)算法存在的阈值更新不够灵活,难以消除鬼影,对阴影较为敏感的问题,提出一种改进的算法。采用非线性阈值调整机制,根据背景复杂度及时调整阈值;对像素被判定为前景点的次数进行控制,去除难以消失的鬼影点;对判定为前景的区域边界点进行二次对比,减少鬼影点;引入基于颜色不变量的阴影检测去除前景中的阴影区域,提高分割精度。在CDnet2014(ChangeDetection.net 2014)数据集上的测试结果表明,改进后的方法相比PBAS分割精度更高,能够很好地去除鬼影和阴影。

In view of the problems of PBAS,which are inflexible in threshold updating,difficult in handling ghost artifacts and sensitive to shadow,an improved approach was presented.Nonlinear threshold adjustment mechanism was applied to adjust the threshold timely according to the background complexity.The time of pixels classified as foreground was controlled to eliminate stubborn ghost artifacts.A second comparison was carried out for boundary points of foreground to reduce the ghost artifacts points.The shadow detection based on invariant color feature was introduced to tackle the shadow in foreground regions and improve the accuracy.The evaluation on CDnet2014 datasets shows that the improved method has superior accuracy over PBAS and it can handle the ghost artifacts and shadow.