A method for determining the likelihood that a pixel in a video image sequence represents true motion in the image. Light from a scene is captured to yield an original image having a sequence of frames. A prototype image is created from the original image as a geometrically accurate representation of the scene which is substantially free of contamination. A tilt variance-Gaussian mixture model and/or multi-variate Gaussian model, based upon optical turbulence statistics associated with the scene, is developed. Each pixel in the original image sequence of frames is evaluated against the prototype image using a probability density function, to yield a likelihood that the respective pixel represents true motion.