Segmenting independently moving objects from egomotion flow fields

Pauwels, Karl and Van Hulle, Marc

Proceedings of the Early Cognitive Vision Workshop, Isle of Skye, 2004

We describe a robust algorithm for the estimation of egomotion from noisy optic ﬂow ﬁelds. The method is nonlinear, conceptually simple, and corrects for the well-known bias to which many egomotion-estimation algorithms are prone. Furthermore, techniques from robust statistics are employed to render the algorithm insensitive to the presence of large numbers of highly correlated outliers. These properties make the algorithm particularly suitable for the segmentation of independently moving objects. The motion ﬁelds of these objects are not an effect of the observer’s motion (as is the case with the static environment), but of the motion of the objects themselves. We demonstrate the procedure on real world image sequences, recorded during car driving situations.