Optical flow field segmentation in an omnidirectional camera image based on known camera motion

摘要

We consider the problem of optical flow analysis in an omnidirectional camera image, in order to segment out dynamic objects from ego-motion caused optical flow. The aim is to use the camera on a mobile robot, thus making it reasonable to assume knowledge of the camera movement in the horizontal plane. Estimation of the optical flow is performed in the omnidirectional image. By knowing the ego-motion, it is possible to calculate the resulting optical flow field with the necessary assumption of planarity of the surrounding scene (the mobile robot is in an open, flat space). However, comparing the calculated and estimated flow field, with the latter assumption by comparing the flow vectors or winding, did not result in satisfactory estimates. Therefore, we propose to geometrically map the optical flow vectors on a sphere, centered at the center of the central panoramic camera projection, since we observed the space angle to be invariant to the errors of the planar space assumption. In the experiments we show that the proposed approach is more robust than comparing the flows in the image plane. In addition, properties of the observed angle for the inverse problem of determining camera ego-motion are considered.