Motion based vehicle detection on motorways

In this paper I describe a vision based algorithm for the detection of vehicles, overtaking or getting into lane on motorways. The transversal components of the image flow field are used. The basic idea is to extract the parts of the motion field which do not obey the epipolar constraint [1], assuming a translatory or a combination of a translatory and a rotatory ego-motion field. In the first case the induced motion field radially diverges from the FOE (focus of expansion) and all nonradial components of the motion field belong to objects with a relative movement related to the observer. This relative motion (e.g. of an overtaking car with the motion vectors pointing in the opposite direction of the induced motion field) is independent from the ego-motion field and easily can be detected and predicted. A feature-based optical flow approach is used, permitting fast computation for the estimation of the image flow. A preprocessing, consisting of a local orientation coding, extracts the image features relevant for optical flow computation and encodes them in bit strings. The following correlation measurement of the feature correspondences reduces therefore to a fast XOR-operation, where the Hamming distance defines the distance between two image features in a spatiotemporal neighborhood. The motion field is calculated for different velocities at different scales of a gaussian pyramid. Then a set of direction-selective prediction filters is applied. The outputs are fuzzy regions of interest at locations where overtaking or merging vehicles can be observed.