Image Registration Algorithm for Deep Learning-Based Stereo Visual Control of Mobile Robots

Abstract

Since the emergence of deep learning as a dominant technique for numerous tasks in the computer vision domain, the robotics community has strived to utilize its potential. Deep learning represents a framework capable of learning the most complex models necessary to carry out various robotic tasks. We propose to integrate deep learning and one of the fundamental robotic algorithms—visual servoing. Fully convolutional neural networks are used for semantic segmentation, which represents the process of labeling every pixel within the image. The obtained information from labeled (categorical) images can be crucial for mobile robot control in dynamic environments. To adequately utilize semantic segmentation for mobile robot control, the segmented images acquired at the desired and the current pose need to be registered (aligned). Since the accuracy of visual servoing depends on the accuracy of the image registration process, we propose to increase the accuracy of mobile robot positioning by analyzing three different optimization algorithms devoted to the registration of categorical images. The standard gradient descent algorithm is compared to the OnePlusOneEvolutionary algorithm, and simulated annealing. Moreover, different cost functions such as Mattes mutual information, global accuracy, and mean intersection over union are also investigated. All the algorithms are tested on our own wheeled mobile robot RAICO (Robot with Artificial Intelligence based COgnition) developed within the Laboratory for robotics and artificial intelligence. The results indicate that the algorithm with a larger exploration to exploitation ratio provides better results. Moreover, the cost function with the steepest convex domain is more advantageous.