Research On The Extraction Of Salient Object Based On Spatio-Temporal Analysis

Salient object extraction is a significant topic in the field of computer vision research because it can detect significant regions(object)with independent visual significance in visual input,which is of great significance for advanced visual tasks such as follow-up behavior analysis,semantic judgment and scene understanding.In recent years,with the extensive application of wearable video equipment,salient object extraction based on Egocentric Video has attracted wide attention of researchers.Based on the spatio-temporal analysis method,it is of great significance to study the extraction of salient object based on the first-view video for the characteristics of Egocentric Video.In this paper,a number of different perspectives,first-view image sequence captured the same time as an input,combined with spatio-temporal analysis.The significance detection algorithm and the target detection algorithm based on the first perspective video are studied.The main work of this paper is as follows:Firstly,we study the fusion significance algorithm based on spatio-temporal analysis.Aiming at the characteristics of the background change of the first-view video,the difference of the object scale and the strong time-varying angle of view,a method combining the bottom-up image salience,the ego-motion information and the object motion information is used to calculate the saliency of the object in first-view video.The salient test results show that the proposed method is superior to IT,GB,SR,CA and AGV in detecting the salience of the first view video.Secondly,we study algorithm of salient object extraction based on fusion significance.A Canny edge detection method is proposed to reduce the number of edges in the background.According to the closeness of the Gestalt Rule,the closed relation between groups is obtained.By analyzing the closeness relation between the composition elements and the salience of the first view video,the objective function of measuring the closedness of grouping is proposed.And the minimum weight optimal matching algorithm is used to solve the objective function to obtain the closed contour of the final object.Experiments show that the proposed algorithm can obtain higher quality moving object contours in the first view video compared with the RRC contour grouping algorithm.Finally,we study algorithm of significant object extraction based on multi-view spatio-temporal analysis.And a method of significant object extraction based on group detection is proposed.Aiming at the first view video and combining the spatio-temporal information,the method of instant attention and spatio-temporal attention based on attention clue is proposed,and the relationship between objects is measured.The algorithm of group detection based on adaptive K-Means clustering is proposed.Group test results,the final can be drawn object(photographers)are most concerned about the object.The average correctness of the algorithm in the Party Scene dataset is 92.5%.