Efficient background subtraction under abrupt illumination variations

摘要: Background subtraction techniques require high segmentation quality and low computational cost. Achieving accuracy is difficult under abrupt illumination changes. We develop a new background method in an expectation maximization (EM) framework. describe foreground colors ratios using few Gaussian mixture models. EM convergence dependent on its initialization. propose novel initialization that considers reflectance implicitly. Scene points occluded by object tend to have prominent since both the are different. introduce topological approach based Morse theory pre-classify pixels into background. Moreover, we only decompose probability distributions initial step our EM. Later iterations do not consider distribution decomposition anymore. The experimental results demonstrate formulation provides variations illumination. Additionally, comparison with one of state-of-the-art methods EM, converges fewer iterations, yielding savings.