Optical Flow-Based Place Recognition: Bridging the Gap Between Simulation and Real-World Experiments

In this research a non-conventional approach is used for modeling an a priori unknown environment. Rather than taking into consideration the properties that describe an environment such as shape, size, or color, we model an unknown environment, with the view for a robot to recognize places, upon re-visiting, using the properties of camera motion. Furthermore, we make use of only a single camera sensor and no variables are known during the place recognition phase. In particular, we model the optical flow vectors magnitudes as a function of velocity and distance. Based on this observation, our algorithm is trained with a large number of varying velocities and distances, and a probability density function is inferred which expresses the relationship between velocity, distance, and optical flow vectors. Moreover, we try to fill the gap between simulated and real environments by directly utilizing the training inferences from simulation to the testing phase in real environments. Our model is evaluated in a real-world environment using a mobile robot that recognizes places from optical flow patterns. The results presented in this paper validate the feasibility of our algorithm.