Compressão de nuvens de pontos dinâmicas: uma abordagem eficiente para a etapa de predição

Abstract

Currently, new technologies for representing scenes with greater realism are the subject of intense research. In this context, 3D multimedia content is gaining focus and enabling applications such as tele-immersive videos, virtual, augmented and mixed reality, mapping of historical objects and architectural monuments, among other possibilities. Point clouds have attracted the attention of current research as a way of representing 3D content because of its low computational cost considering aspects such as data generation and manipulation, thus standing out against alternatives such as meshes. However, as in 2D digital videos, the volume of data generated by point clouds without compression becomes prohibitive for transmission and storage. In the video compression scenario, redundancies are explored, including spatial redundancy, called intra-frame prediction and temporal redundancy, called inter-frame. The temporal redundancy present in the videos, for example, allows the most significant gains in compression. However, characteristics inherent to the point cloud as different boundaries of the bounding boxes between temporally neighboring clouds, as well as the presence of occluded or sparse regions, makes it difficult to explore these predictions in compression. In this sense, dynamic point cloud compression, especially the intra and inter-cloud prediction stages, is an active and challenging topic, still little explored in the literature. The works found in the literature that deal with inter-cloud prediction through macroblocks, for example, explore only fixed size and/or explore only colocalized macroblocks. More recent proposals also suggest planning point clouds and using video encoders. However, with this approach, the exploration of three-dimensional space ends up being neglected in the inter-cloud prediction stage. This thesis aims to present an efficient approach to predict the compression of dynamic point clouds. The proposed solution includes a motion estimation based on macroblocks in 3D space, in addition to the ability to operate on different block size configurations, as well as a mode decision algorithm to choose the best configuration based on the relationship between bit rate and quality. The solution was implemented and evaluated following the common test conditions proposed by MPEG, which include the evaluation of five dynamic point cloud sequences, each with approximately 300 point clouds. The rate and quality results were compared with those obtained in the anchor software for dynamic point cloud compression. The approaches developed show significant gains, both in terms of quality and in compression rate.