Perception-driven approaches to real-time remote immersive visualization

In remote immersive visualization systems, real-time 3D perception through RGB-D cameras, combined with modern Virtual Reality (VR) interfaces, enhances the user’s sense of presence in a remote scene through 3D reconstruction rendered in a remote immersive visualization system. Particularly, in situations when there is a need to visualize, explore and perform tasks in inaccessible environments, too hazardous or distant. However, a remote visualization system requires the entire pipeline from 3D data acquisition to VR rendering satisfies the speed, throughput, and high visual realism. Mainly when using point-cloud, there is a fundamental quality difference between the acquired data of the physical world and the displayed data because of network latency and throughput limitations that negatively impact the sense of presence and provoke cybersickness. This thesis presents state-of-the-art research to address these problems by taking the human visual system as inspiration, from sensor data acquisition to VR rendering. The human visual system does not have a uniform vision across the field of view; It has the sharpest visual acuity at the center of the field of view. The acuity falls off towards the periphery. The peripheral vision provides lower resolution to guide the eye movements so that the central vision visits all the interesting crucial parts. As a first contribution, the thesis developed remote visualization strategies that utilize the acuity fall-off to facilitate the processing, transmission, buffering, and rendering in VR of 3D reconstructed scenes while simultaneously reducing throughput requirements and latency. As a second contribution, the thesis looked into attentional mechanisms to select and draw user engagement to specific information from the dynamic spatio-temporal environment. It proposed a strategy to analyze the remote scene concerning the 3D structure of the scene, its layout, and the spatial, functional, and semantic relationships between objects in the scene. The strategy primarily focuses on analyzing the scene with models the human visual perception uses. It sets a more significant proportion of computational resources on objects of interest and creates a more realistic visualization. As a supplementary contribution, A new volumetric point-cloud density-based Peak Signal-to-Noise Ratio (PSNR) metric is proposed to evaluate the introduced techniques. An in-depth evaluation of the presented systems, comparative examination of the proposed point cloud metric, user studies, and experiments demonstrated that the methods introduced in this thesis are visually superior while significantly reducing latency and throughput.