Reconstrução tridimensional com múltiplas câmeras e sistema distribuído sob o paradigma fog

Abstract

Whether being for generating electronic games or tools of assistance to professional operators, the study and development of virtual and augmented reality systems has reached great proportions. As ways of reproducing three-dimensional objects and scenarios in virtual environments, there is reconstruction software as a faster and less costly alternative in specialized labor when compared with the standard form of generating these objects, through graphic design with specific software. There are several ways to perform 3D reconstruction presented in the literature, each with its characteristics and peculiarities, such as stereo vision, monocular SLAM, or even with RGB-D cameras. Therefore, in this work is proposed a mixed approach, using advantages of different methods, such as the accurate odometry from the stereo cameras allied with the well-defined map from the RGB-D sensor, through a distributed FOG system. The communication between processes and machines is done through the publisher-subscriber model, using the framework ROS for the message exchange. This composition allows the division of the computational costs, as well as increasing the scalability, which makes possible that new sensors and methods could be added in the future in an easy way. The obtained 3D models, in point cloud format, were comparable to the presented by the best software within the literature, mainly when used for large objects or environments reconstruction. Moreover, the stereo vision applied to obtain the odometry has shown to be more effective in the tested paths concerning the point clouds, justifying the application with multiple sensors. The odometry error, when compared to the robotic arm as the ground truth, was approximately 5 millimeters in the worst result. The implementation of the distributed system was realized successfully, having, however, a greater processing time due to the message exchange through a LAN network. The obtained results suggest the possibility of applying the acquisition step with embedded systems. Through these clouds, it is possible to apply mesh generation algorithms, obtaining the objects or scenarios to be used in virtual or augmented reality applications.