Abstract
A remaining challenge in autonomous mobile robotics is navigation in unknown environments. While methods such as Simultaneous Localization and Mapping (SLAM) address this problem, they can be computationally demanding for small embedded systems. An alternative approach is inspired by biological navigation strategies, such as optical flow–based navigation observed in bees. This thesis investigates whether such a strategy can be implemented on a small mobile robot equipped with a catadioptric vision system and a Raspberry Pi as the onboard computer. The robot was fitted with a custom-designed catadioptric setup providing full 360° horizontal perception. A computationally efficient pipeline combining empirical panoramic unwarping, dense Farnebäck optical flow estimation, and divergence based Time-To-Contact (TTC) calculation was implemented under embedded hardware constraints. The system was evaluated in random-walk obstacle avoidance experiments in an arena with varying numbers of obstacles, as well as in a tunnel setup to assess whether the robot exhibits centering behavior similar to that observed in insects. The results show improved performance compared to previous experiments conducted on the same platform using a single forward-facing monocular camera, particularly by reducing blind spots. How ever, performance degradation was observed in low-texture environments, under non-uniform lighting conditions, and as obstacle density increased. These findings suggest that catadioptric optical flow–based navigation is feasible on resource-constrained platforms, although further improvements in calibration and controller design are required to achieve robust performance.
About the speaker
Vassil Radoynovski is a Bachelor’s student at the University of Konstanz with an interest in computer vision and bio-inspired robotics. He is currently working under the supervision of Dr. Andreagiovanni Reina on optical-flow-based navigation. In his ongoing research, he focuses on implementing omnidirectional vision for reactive navigation on the Ringattractor.