Implementation of Vertical-Rectification and CNN Models for an Analogic Range-Estimation Algorithm from a Stream of Images

The implementation of autonomous mobile robots in real life environments still has numerous challenges to face. The most crucial problem is real-time decision-making, using appropriate methods with the right hardware. Recovering the three-dimension scene geometry and detecting moving targets simultaneously from a stream of images are important tasks and have wide applicability in the creation of autonomous mobile robots, such as persistent choice of a safe route free of obstacles, targeting objects to avoid collisions, autonomous navigation and robot manipulation. In the present work, we focus on exploiting the robustness of the analogic-array-processing-aspect introduced by the Cellular Nonlinear Network paradigm to develop a real time tracking method for a stream of general signals coming from space-distributed sources for monocular autonomous mobile robots. The motivation for developing the new tracking method is from one hand the matching operation has to be performed in real-time, while from the other hand a 32 bit floating point accuracy is not often required, which, together with a vertical rectification, as an intermediate process to minimize the token relative displacements between two frames, can lead to a robust real-time object tracking system. The technique has been successfully applied to several indoor sequences of images. The results of the simulations are presented and discussed.