Visual Saliency Algorithm Research On 3D Dynamic Model In Spatio-temporal Domain

Thanks to the rapid evolution of computer hardware devices,3D technology has been applied to many fields,including entertainment,biology,medicine,education,military,etc.The impact of the 3D technology to people's daily life is compound.One of the key characteristics of primates is the capability of analyzing an observed complex scene autonomously at short notice.The salient region represents the area in which the human vison is highly sensitive,therefore saliency detection technology allows the computer vision system to emulate visual mechanism of human eyes and obtain the area of interest to human eyes.Saliency research on 3D scene could filter out those redundant information for saving computational resources.Traditional 2D image saliency detection technology has been developed for several decades and there are many classical models as theoretical basis.However,the complete metrics for calculating the significance of saliency in 3D scene is lacked.What's more,the saliency analysis is even more complicated as the 3D scene that people observing is often dynamic.This thesis proposed a time-spatial domain visual saliency detection algorithm for dynamic 3D scenes.1)For each frame,the Mesh Vertex Curvature algorithm with Guassian kernel filter and the Center-Surround Differentiation algorithm are applied together to calculate the mesh saliency.2)Different saliency weights are assigned to corresponding area of the virtual scenes based on the distances to the observer which can highlighting the area closer to the observer.After that a saliency map in spatial domain is obtained.3)The motion information between frames is calculated by comparing one frame with all the other frames in one specific time period.The results are further optimized according to the differences of local motions,the continuity of the motion states and the persistence of vision.After that a saliency map of the 3D mesh model in the time domain is obtained.4)Finally,the 3D saliency map in time-space domain is obtained by merging both saliency maps in time domain and spatial domain.At length in this thesis,Swarm Method and Multidimensional Characteristics Comparison Method are used to compare the results of the saliency detection algorithm in terms of every subject's gaze points at each frame and gaze plots of entire observation process respectively.Multi-dimensional Characteristics Comparison Method is an innovative multi-dimensional scanpath similarity metric in 3D scene based on path matching algorithm in 2D.A human experience test is conducted to compare the measured gaze data with the saliency detection algorithm.The experience verifies the variety of the proposed algorithm.