Obstacle detection by real-time optical flow evaluation

Reliable estimation of optical flow vectors facilitates the evaluation of the motion of a camera relative to the environment. Furthermore, optical flow vectors comprise information about the 3D structure of the recorded scene. In this contribution we present an approach for the detection of stationary obstacles and moving objects in front of a vehicle. The detection of obstacles and other objects is based on the evaluation of optical flow. The optical flow vectors are calculated with a local analytical approach. The necessary first and second order spatio-temporal derivatives are computed in real-time with our second-generation custom designed image sequence analysis system MiniVISTA. At all image locations it is tested whether the structure of the gray value distribution is sufficient for a local optical flow vector calculation or not. Only those sets of spatio-temporal derivatives which passed this structure test are transmitted to subsequent processes being executed on a transputer network. Experimental results obtained from image sequences recorded on our experimental vehicles MB 609D and BMW 735 iL are presented.