A System To Capture And Render Dynamic Light Field Information Based On Distributed Stream Computing Architecture

Three-dimensional scene information collection, capture and model on the targetobject has been one of the hot research in computer vision and computer graphics, thepurpose is to let the computer like human to perceive scenes, obtain information, beable to accurately and quickly reproduce and restore scene. In order to obtain moreaccurate three-dimensional model, people need to collect more data, higherdimensional light field information, the algorithm is more complex which makesrendering system to put forward higher requirements in computing performance, theincreasingmass of data is becoming abottleneck in three-dimensional rendering，particularly in the real-time rendering requirements occasions.This paper focuses on the issues that how acquire dynamic light field data andhow to render in real-time. Research as follows,1) How to control the camera tosync collect and storedata in cluster，we achieve a distributed real-time acquisitionsystem.2) How to use a distributed platform to handle massive dynamic light fielddata, we achieve a batch-based mass data processing system.3) How to process andrender the massive dynamic light field data in real-time, we achieve a distributedstreaming real-time processing system.The first chapter, introducing thecurrent situationof dynamic light fieldacquisition and reconstruction, as well as paper handling problems encounteredmassive data, and then introduced the development and status of the framework of bigdata and the application in our system.The second chapter describes the design and architecture of distributed real-timedynamic light field acquisition systems. The system is strong and flexible, and easy toexpand. By operating the remote control command in master node to control thecluster,and fine-grained to control the cameras’ startup, acquisition and closed; bydatabase to manage the storage location of collected data in the cluster, greatlyreducing the storage burden on the master node; to synchronous acquisition, thesystem also sets the time server, ensure consistency on the acquisition time; and thesystem has played a satisfactory performance.The third chapter describes a distributed processing system which handles lightfield data in batch mode,and3D modeling algorithm we used is IBVH algorithmwhich is simple and rapid. In this paper, we use big data processing platforms like spark and hadoop to reconstruct the target object in the scene, implements a custompicture data types and some corresponding three-dimensional point cloud input andoutput interfaces to meet the input and output of the platforms.The fourth chapter describes a distributed processing system which handles lightfield data in real-time mode. Data flow as follows, acquisition command issued fromthe master node, the cameras capture and transmit data to the relay node, and thenwrite to HDFS, SparkStreaming cluster gets the data and starts to processing, andoutput the final results of the three-dimensional point cloud. For different operatingsystem, once data is collected and it will do sometransformation to meet the needsboast platform;to ensure data integrity in acquisition and transmission, we check thedata both in sender and receiver. Because of the limitation of hardware networkdevice, we do some control the acquisition rate; for the delay, weoptimize the systemand try to balance the input data in a unit time to match the cluster performance.The last chapter summarizes the research work and the prospects for futureresearch dissertation work.