基于鲁棒M估计和Mean Shift聚类的动态背景下运动目标检测

Detection of Moving Objects in Dynamic Scenes Based on Robust M-estimator and Mean Shift Clustering

针对动态背景下运动目标的检测问题,提出了一种基于鲁棒M估计和Mean Shift聚类的目标检测新方法。首先,在考虑全局光照变化的情况下,构建鲁棒M估计器估计全局运动,以实现最小化相邻2帧图像中所有像素亮度的绝对残差和,根据M估计得到像素点权值,提取出代表局部运动信息的离群点;在离群点中均匀抽取网格点,然后利用Mean Shift聚类算法实现不同运动点的分割;根据聚类的结果生成凸包,准确分割出运动目标区域。实验结果表明,该方法能检测出动态背景下的多个运动目标区域,多目标检测准确度到达95%以上,并且只需两帧图像就可以准确检测并锁定运动目标,满足实时处理的要求,具有一定的工程意义.

Focusing on the problem of moving objects detection in dynamic scenes, a novel algorithm based on robust M-estimator and mean shift clustering was proposed. First, considering the case of global illumination change, M-estimator was constructed to estimate the global motion in order to minimize the absolute residuals of pixels luminance between two adjacent frames. The structured outliers could be extracted according to the weight map of every pixel. Then the grid points were selected evenly from outliers and different point belong different moving object was clustered by mean shift algorithm. The convex hulls were generated under clustering results, to accurately segment the moving object regions. Experimental results show that this method can accurately detect multiple moving objects in dynamic scenes, and MODA can reach 95%. Besides, only two frames are needed to detect and lock the moving objects by this algorithm, which can meet real-time processing requirements and has a certain degree of engineering significance.