Detection of moving objects in dynamic scenes based on robust M-estimator and mean shift clustering

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.