Research On Visual Object Tracking Based On Saliency

As a fundamental research topic of computer vision,visual tracking has a wide range of application,such as video monitor,human computer interaction and autonomous navigation.It has aroused increasing interest in both academic and industry.But,it is still difficult to design a robust tracker due to the presence of various challenges,such as shape deformation,illumination variation,scale changes,occlusion and background clutter.The study on psychology reveals that a mechanism of selective attention exists in our human visual system.It makes us have an excellent ability to quickly catch salient information from complex scenes.This selective attention mechanism can great promote the efficiency of data processing.We first investigates the problem of video saliency detection.Then,the object appearance modeling and data association for visual tracking are investigated.Finally,we combine the saliency detection with visual tracking to promote the tracking performance.The detailed research contents and main contributions of the thesis are as follows:1.we investigate the problem of spatiotemporal saliency detection based on the bottom-up visual stimuli features.An spatiotemporal saliency detection method via global manner is proposed.Given a video,the method first computes the spatial saliency and temporal saliency respectively.Then,motion entropy is utilized to fuse the two saliency maps adaptively.For computing spatial saliency,three factors are considered:spatial constraint,color double-opponent,and similarity distribution.As for the temporal saliency,we compute the global motion contrast of dense optical flow.To suppress the motion noise,a new histogram of average optical flow(HOAOF)is proposed to compute the motion contrast of different pixels.In contrast to conventional video saliency detection methods,the proposed mothed can obtian superior performance.2.The study on psychology reveals that high level features in a scene play an important role in attracting human visual attention.Based on the psychology theory,we propose a video saliency detection method by integrating bottom-up and top-down visual stimulus.Take the news video as an example,we investigate the influence of high level features in special video when detecting saliency.In the bottom-up attention model,we use quaternion discrete cosine transform in multi-scale and multiple color spaces to detect static saliency.Meanwhile,multiscale local motion and global motion conspicuity maps are computed and integrated into motion saliency map.In the top-down attention model,we utilize high level stimulus in news video,such as face,person,car,speaker,and flash,to generate the top-down saliency map.By considering high level features,the proposed method can obtain better result.3.To deal with occlusion,rotation,and scale changes,we propose a visual tracking method via consistent features selection.The consistence of feature is defined by geometry constraint,and then selected by a revisied density clustering method for object tracking.Furthermore,we utilize color histrogram as object's reference model which is used to meature the confidence of tracking result.According the computed confidence,the object's appearance model is updated.The proposed method can deal with object's rotation,scale changes and partially occlusion.Even when object suffers from fully occlusion,our method have a chance to recover the object.4.Since it is difficult to detect and match features in degraded image and plain image region,we propose a visual tracking method via consistent discriminative region.The method utilize exemplar-SVM to select discriminative regions which are used to model object's appearance.Then,the selected regions are tracked by correlation filters respectively.The region's consistence is defined by two factors: trackable likehood and local predictive power.According to the computed region's consistence,the object's position is determined globally by voting scheme.Meanwhile,the object's appearance model is updated adaptively.The experiments show that the proposed method can promote the tracking performance.5.According to our study in previous,we propose a visual tracking method via weighted salient features.The object appearance is modeled by two statistical models.One is the structure preserving statistical appearance model.Another is an adaptive salient feature statistical appearance model.During the construction of adaptive salient feature statistical appearance model,the discriminative salient features are selected online automatically.The object is tracked in particle filter framework.The particle weight is computed by fusing two appearance model's likelihood which is computed by comparing paticle's two appearance models with reference models.The fusion weights is adjusted automatically according to the computed likelihoods.Experiments show that the tracking performance is promoted by introducing salient appearance features.