Visual Deep Learning-Based Mobile Robot Control: A Novel Weighted Fitness Function-Based Image Registration Model

The recent development of faster and more accurate deep learning models has enabled researchers to utilize the potential of deep learning in robotics. Convolutional neural networks used for the process of semantic segmentation are being applied to improve the traditional robotic tasks by adding an additional level of intelligence, through the execution of context-aware tasks. Having that in mind, visual servoing can now be performed in a completely new manner, by exploiting only semantic and geometric knowledge about the environment. To carry out visual servoing, the mathematical model of the error between the images generated at the current and the desired mobile robot pose (i.e. position and orientation) in the image space needs to be adequately defined. In this paper, we propose the novel mathematical model for the weighted fitness function evaluation, which is utilized for the image registration process within the visual servoing framework. By weighting the classes by their importance in the desired image, the convergence domain of the initial error in the visual servoing process can be greatly extended. The experimental evaluation is carried out on the mobile robot RAICO (Robot with Artificial Intelligence based COgnition), where it is shown that weighted fitness function enables more robust intelligent visual servoing systems with a lower possibility of failure, easier real-world implementation, and feasible object driven navigation.