Reactive avoidance of dynamic obstacles through optimization of their epipoles

This paper proposes a new reactive method of avoiding dynamic obstacles by real-time optimization. A mobile robot is equipped with a monocular camera and dynamic obstacles are identified in the image sequence by clustering the optical flow. The respective epipoles of the clusters are determined and afterwards the relative epipole positions are evaluated to identify colliding and dangerous objects. In this work the correlation between the vehicle's velocities and the cluster epipoles is derived and utilized in the proposed cost function for shifting the epipoles. By optimizing this cost function the vehicle is able to avoid collisions, which means that the 3D motion is deduced from purely 2D image data. Finally, the validity of the concept is confirmed by hardware-in-the-loop simulations.