ABSTRACT
Classical motion-detection methods, such as those presented by Stauffer and Grimson [7] or Elgammal et al. [3], use statistical temporal modeling for each background pixel. Afterwards, optional post-processing steps are applied with the aim to introduce spatial consideration such as object compactness. Moreover, most approaches from the literature only use a background model. In such approaches, motion detection is performed independently for each pixel, by evaluating whether each pixel is sufficiently similar to the background model. Furthermore, when the background and the target (foreground) appearance are close (similar colors, for instance), it becomes difficult to discriminate them without a priori knowledge. If we want an efficient and generic method, this information can then only be obtained with an updatable target model in addition to the background one. By looking for the closest model for each new incoming data, the detection process is therefore much more accurate.
