Background subtraction for vehicle detection

We propose a background subtraction method for segmenting vehicles from background in a video sequence. Our method can be considered as a hybrid of two existing methods. The background is modelled per-pixel with a collection of pixel values. The foreground/background decision is based on whether the current pixel value finds a match with the samples in the model. The matching samples in the model are adapted towards the current pixel value with a learning parameter. Evaluation tests on the public CDnet dataset shows good results in a highway scenario.