Study On Object Of Interest Extraction And Tracking Methods Based On Visual Attention Mechanism

Due to the complexity of the real-world application scenes,such as illumination change,the influence of various clutter and noise in the background,the interference of various uncertain factors in the outdoor scene,as well as the variation of pose,distortion of shape,change of size,and occlusion,the object of interest extraction and tracking has become a commonly recognized challenging issue in computer vision domain.Currently,most of the proposed algorithms have good performance in solving one or several problems under specific conditions or applications,however,there is still lack of methods that can solve all kinds of uncertain conditions and adapt to different applications.As we all know,human can distinguish foreground from background and recognize objects effortlessly no matter how the scene changes.Hence,in depth analysis and study on object of interest extraction and tracking based on the mechanism of human visual attention has crucial theoretical and practical values.The main work of this dissertation is as follows:1)In-depth theoretical study and research on the biological basis of human visual system,human perception and visual attention mechanism is made with the combination of multi-disciplinary research results,such as neurobiology,psychology and cognitive science.The current representative psychological models and computational models of visual attention mechanism are summarized and analyzed comprehensively.2)According to the problem that the initial contour or mask is given manually in the majority of current active contour algorithms,an improved active contour without edges algorithm is put forward for image segmentation based on selective attention.The proposed algorithm mainly contains three parts: image preprocessing,mask initialization and segmentation based on evolution.Some pixel statistics such as the positive detection rate,the false detection rate and compare convergence time and so on are used to quantify the results of the experiment.Experimental results show that the proposed algorithm not only can take place of the traditional methods which select the initial mask manually to realize the segmentation completely automatic,but also can get more accurate segmentation results for multi-object image.3)As to the salient object extraction of traditional visual attention methods are not accurate enough,a new algorithm is poposed based on fast dictionary learning and feature rarity for salient object extraction in natural images.Firstly,fast dictionary learning is applied to feature extraction based on sparse coding algorithm,then the saliency is computed by feature rarity and the restriction of sparse coefficients.Finally,the mathematical morphology operators are used to remove false objects.At the same time,the ROC curve is used to quantify the experimental results for comparison.Experimental results show that the proposed method is more accurate for extracting the salient object compared with the traditional techniques and can also deal with images containing multiple objects.4)To deal with the problem of manual selection of object or background modeling of traditional object tracking algorithms,inspired by the dynamic human visual attention mechanism,an object tracking algorithm based on dynamic visual saliency and multi-feature particle filter is put forward.Firstly,the dynamic visual saliency is modeled by SIFT flow and the object of interest to be tracked is detected by the combination of dynamic saliency and static saliency.And then the selective attention and multi-feature particle filter are combined to realize the detection and tracking of the object of interest.Experimental results show that the proposed method not only can extract the object of interest effectively,but also can track the object of interest stablly.5)Due to most of the current object tracking algorithms will fail when serious occlusions occur,inspired by the human perception memory mechanism and visual attention mechanism,an object tracking algorithm combined with visual saliency and the spinning tri-layer-circle memory model is proposed.First,the object of interest is extracted automatically based on visual saliency.Then the spinning tri-layer-circle memory model imitating the human memory information storage and extraction process is constructed and applied to the updating of the object template in moving object tracking process.Finally,the spinning tri-layer-circle memory model is incorporated into the particle filter algorithm to complete object of interest tracking.Experimental results show that the proposed algorithm not only can extract the object of interest effectively and automatically,but also can track the object of interest when serious occlusions occur,and the performance is better than the htraditional object of interest tracking algorithms.