OBSTACLE DETECTION BY EVALUATION OF OPTICAL-FLOW FIELDS FROM IMAGE SEQUENCES

In comparison with single images, image sequences contain information about the dynamic aspects of the recorded scene. Dynamic aspects play a crucial role in traffic scenarios, especially for the guidance of an autonomous vehicle. In this contribution we present an approach for the detection of stationary obstacles and moving objects on the road of an autonomous vehicle by evaluation of optical flow fields from image sequences. It will be shown how two-dimensional optical flow fields can be interpreted to infer information about the three-dimensional environment if the camera is moving on a planar surface. Objects are detected by comparing calculated optical flow fields with a model vector field which explicitly expresses the image shift expectation resulting from the camera motion if the environment is free of obstacles. Until now, we had to determine the camera motion heuristically in order to estimate a model vector field. In this paper we evaluate additional velocity information from vehicle sensors. It will be shown that the obstacle detection procedure using velocity information from vehicle sensors yields results similar to the ones obtained by the heuristic velocity estimation. Results obtained from investigations with three different approaches for the estimation of optical flow vectors are used to test the obstacle algorithm.