Computational Modeling Of Visual Saliency And Its Application In Extraction And Tracking Of The Objects Of Interest

The extraction and tracking of objects of interest from an image or a video sequence is an important research content in the field of computer vision and has wide application prospect.However,there may be uncertain factors such as noise,interference and illumination change in different specific applications,and the targets may also have changes in deformation,scaling and motion speed,which will adversely affect the accuracy of target detection and tracking.Therefore,how to identify,extract and track objects of interest in complex scenes has always been a challenging subject in this research field.At present,most of the existing algorithms are based on a certain specific situation or scene to achieve better performance by overcoming one or more of the above disadvantages,while there are few algorithms that can simultaneously solve all the disadvantages in any scene.In contrast,the human visual system has the ability to“detect” interesting targets from various scenes with no effort required.Therefore,it is of great theoretical and practical significance to study the biological mechanism of visual attention and calculate visual saliency based on it,and to explore the extraction and tracking of objects of interest.The main work of this dissertation is as follows:1.Based on the related research results of psychology,cognitive neuroscience and anatomy,the human visual saliency mechanism is studied.The typical psychological models and computational models of visual saliency are analyzed and summarized in depth.2.In view of the fact that the traditional detection algorithms are not accurate enough to extract the objects of interest in complex scenes and may have high computational cost,and it is particularly important to get "the most significant" location or region in most practical detection applications.Inspired by the neurobiological framework,a "two-stage" computing algorithm for saliency acquisition is proposed based on information gain.In the first stage,a rough pre-attention region is obtained through the selected feature points,and in the second stage the saliency at each position is extracted by the computation of information gain over the pre-attention area.The advantage of this algorithm is that it can be adjusted flexibly between computational cost and extraction precision according to the different requirements of accuracies in different application situations.3.According to the problem that it is difficult to accurately extract moving targets in moving scenes,a visual saliency aided fast moving target detection method is proposed.First,an improved random sample consensus algorithm based on memory windows is proposed,which is inspired by the instantaneous memory mechanism of the human visual system.The memory windows are applied in the process of weight calculation during the propagation modeling of the data points.In addition,an importance sampling process of the matched data points is carried out according to the size of their confidences,which leading to an improved convergence speed.To locate the moving foreground targets as soon as possible,the visual saliency is integrated into the Vi Be framework,meanwhile it is used to assist the generation of appropriate dynamic background updating factor and suppress the "ghost" effect.4.In a video sequence,the tracking accuracy may be affected by the possible occlusion,deformation of the targets,and also the random change of the interferences in the scene.To solve these problems,an automatic target tracking algorithm based on Oriented FAST and Rotated BRIEF(ORB)feature points is proposed.The motion edge saliency is extracted from the inter-frame difference and the Lucas-Kanada(LK)optical flow,the edge of each moving target is extracted by the motion edge growth algorithm,and then the target group is separated from the background.During the tracking,the locating and searching of targets are realized by matching ORB feature template.Experimental results show that the proposed algorithm can effectively deal with the disadvantages such as occlusion and illumination changes,and achieves better tracking results than other traditional algorithms.5.The existing human fixation models is not high in distribution accuracy and easily affected by many unfavorable factors.Inspired by the movement trends(Bias)of human eyes when people freely observe the outside scenes,a prediction model of fixation motion based on visual significance is proposed.The model is divided into two parallel processes: global transfer and local shift.The former can predicts the long distance migrations of saccades in the scene by combining the visual saliency,visual centrality and inhibition of return mechanism,while the latter is used to determine the next shift destination by searching the local largest salient point.The experimental results show that the proposed model can effectively predicts and simulates the generation and migration of human fixation,and it has statistical advantages in the distribution of fixations when compare it with some other prediction models under NSS and CC criteria.