A biologically motivated mid-level stage of robust optic flow processing

We test a biologically motivated filtering method [9] for noise decreasing in optical flow fields.Our filter model adopts a property of a particular motion sensitive area of the brain which represents the average incoming motion signals over domains whose size increases with the distance from the center of the projection. We use the task of heading detection from optic flow as a way to estimate improvements of flow fields generated by a standard algorithm. First we test our model on flow fields generated with a modified Nagel-Algorithm from image sequences which are directly calculated from three dimensional real world data assuming a given self motion. Thus we retain the control about the exact heading and rotation and have ground truth. Not surprisingly, due to the noise and the aperture problem the results for the raw flows are often incorrect. In contrast the filtered flows allow correct heading detection. Furthermore we apply our filtering model to flow fields which are calculated from image sequences recorded by a rigid installed camera during car motion. For the unfiltered flow the estimated heading directions differs widely for different subsampled calculations. In contrast, the results obtained from the spacevariant filtered flows are much less variable and therefore more confidential.