Background Subtraction with Outlier Replacement

A data driven background subtraction algorithm where each background pixel is modeled with a representative set of samples is presented. The samples are pixel values observed in preceding frames. Each pixel in an incoming frame is classified as background or foreground by comparing the pixel value with the samples in pixel's background model. The background model is periodically updated by replacing the most outlying sample in the model with the current pixel value. Techniques for suppressing dynamic background and ghost regions are incorporated in the algorithm. Evaluation tests on a public dataset shows markedly faster ghost suppression and fewer false positives in dynamic background regions, as well as improved overall performance in terms of evaluation metrics, compared to the base method.