Real-time Computation Of Depth Information Technology Based On Light Field

Computer vision becomes a hot research direction when the artificial intelligence age appears.Stereo vision is an important subject of computer vision,which has been widely used and developed in many fields such as automatic vehicle,military&space,medical diagnosis,human-computer interaction,AR/VR etc.The most important problem in stereo vision is to extract and reconstruct the depth information of the scene in real time.Therefore,it is very important to research real-time and accurate depth information extraction technology.The light field records the irradiance of all rays in space^[1][2].The 4D light field data,which captured by the light field cameras,contains the spatial and angular information of rays that can be used to extract more depth information of the scene.In this paper,the optical field data is captured by the sparse light field camera array.After the relevant preprocessing,we used the proposed stereo matching algorithm to realize real-time extraction of the scene depth information based on GPU parallel computation on Linux system.After that,we used the disparity map to reconstruct the 3D scene.Firstly,we summarize the related technical background,including the theory of light field,stereo vision system,and GPU parallel computing,etc.,and introduces the latest research progress and application of related technologies.In this paper,the sparse light field data of binocular camera calibration,image is obtained by calibration of the camera's internal and external parameters of polar rectified,and the geometrical imaging principle of stereo vision system and space coordinate relations are analyzed in detail.Then the paper introduces the real-time stereo matching algorithm based on GPU.This paper utilizes CUDA architecture to realize real-time extraction of the scene depth information based on GPU parallel computation.In order to have good robustness in low texture and complex light environment,we chose the Census Transform as matching cost function.The cost aggregation applied SGM with eight path directions;the support window size is 9*7 pixel.Compared with other methods,it has the advantages of smaller amount of calculation and less matching error.WTA?winner take all?strategy was used to chose pixels best disparity value.We have applied left-right consistency check and guided filter for additionaloptimizations,effectively improve the accuracy and efficiency of the stereo matching algorithm.Finally,we used the disparity map to reconstruct the 3D scene.In this paper,30 frames per second?fps?with1920×1080 image resolution and 128 disparity levels using NVIDIA GTX 1080 could be reached.