Visual Object Tracking Method In Complex Scene

One of the main goals of computer vision is to enable computers to have basic functions such as motion perception,object recognition,and scene understanding,just like the human visual system.Visual object tracking is to realize the function of motion perception.It has been one of the hottest research topics in the field of computer vision for the past two decades.Visual object tracking has many practical applications,such as human-computer interaction,video surveillance and traffic scene understanding.Although many object tracking methods have been proposed,it is still a challenging problem to design a robust tracking method due to factors such as partial occlusion,background clutter,illumination and viewpoint changes.The research work in this thesis only focus on monocular visible image,short-term,model-free,single-target and causal tracking,and the main contributions include the following three aspects:(1)For the subspace-learning based tracking methods which ignore background information,we propose a method that tracks object with partial occlusion by background alignment(POBA),aiming to find the best candidate state with accurate occlusion mask.The subspacelearning based tracking methods use a set of basis vectors to reconstruct the candidate state and select the candidate state with minimum reconstruction error as tracking result.It is well known that the accuracy of reconstructed results is seriously affected by partial occlusion.In addition,updating the model with samples containing occluded regions is likely to mislead the tracker to drift.The existing methods either do not consider the partial occlusion problem or only use the current observation and reconstruction result to identify occlusion area.The occluded area usually comes from the background area in previous frame,which is ignored in the existing methods.The POBA tracker treats the current observation as a combination of the object appearance and the occluded area.The appearance of the object is modeled by a set of basis vectors,and it is learned from the gray image by incremental principal component analysis method.Under the assumption that the background is almost the same between two consecutive frames,we generate an occlusion area from the previous frame.In addition,since most candidate states are obviously different from object,they can be filtered by some designed occlusion masks,which can reduce the computational cost.In the experiment,we evaluated the POBA tracker on two challenging datasets.(2)For the online multi-instance learning based tracking methods that update and select weak classifiers without distinction,we propose an online multi-instance learning tracking method with reliable components(OMRC).Online multi-instance learning based methods learn discriminant classifier under the boosting framework.Weak classifiers are learned from partial regions of the object,and all classifiers are updated when the appearance model is updated.However,due to the irregular shape or occlusion of the object,some components do not belong to the object and should not be learned.On the contrary,the discriminative weak classifier learned from these components may cause the tracker to drift.To overcome this problem,we keep a background template and an object template while tracking.By comparing the current tracking result with these two templates,it can be obtained whether the pixel belongs to the object.For a component,the ratio of pixels belonging to the object higher than a predefined threshold is considered to be a reliable component.In addition,in order to better represent the image,we use HOG features and Histogram features to replace the widely used Haar-like features,and design an online learning method to train weak classifier.Experiments were performed on two challenging datasets including OTB2015 and Temple Color.The experimental results prove the robustness of the OMRC tracker and the effectiveness of each part in the OMRC tracker.(3)Considering that the Siamese network based tracking methods cannot adapt to object appearance changes and distinguish similar regions,we propose a tracking method that combining offline learned similarity measurement and online discriminative classification(SMDC).The Siamese network based object tracking methods obtain a similarity measurement through offline training.While tracking,they only rely on the object template in first frame.Sometimes,they get lost due to ignoring object appearance changes and can not effectively distinguish similar regions.In order to capture the information of the sequence during tracking,SMDC adds a linear classification branch.In the tracking process,a small number of similar samples are learned online,which is complementary to the offline learned Siamese network.In addition,the current Siamese network method keeps the aspect ratio of the object unchanged when extracting the object template,and does not use the aspect ratio information of the object template.However,it unreasonable that images with different aspect ratios predict the same candidate region to obtain the corret response and location regression.During the feature extraction process of object template,we add an object template alignment layer to obtain the object template with a uniform aspect ratio.The robustness of the method is verified on multiple datasets.In general,we conduct related research on two challenges in visual object tracking,namely occlusion and similar regions.To solve the problem of occlusion,we construct the background template for the subspace-learning based method and the multi-instance learning based method respectively.In the subspace-learning based tracking method,more accurate reconstruction result can be obtained according to the background information.In the multi-instance learning based tracking method,reliable components are updated and integrated into the final tracker that avoids the interference from background regions.When dealing with similar objects,the Siamese network based method introduces a linear classification branch to exploit intermediate information while tracking,then the tracker can effectively distinguish similar regions.The experimental results of the proposed methods on several public datasets have proved their robustness.