Research On Visual Object Tracking Algorithms Based On Multiple Models In Complex Scenes

Visual object tracking is one of the classic research in the field of computer vision.Its purpose is to estimate the states in the subsequent frames by the image of object in the first frame.It provides the tracklet information for further research such as behavior analysis,anomaly detection,etc.With the continuous development of artificial intelligence and machine learning,visual object tracking has been widely applied in people's lives.Therefore,it not only has high theoretical value,but also has high practical application value.Because of the complexity and diversity of object and scene,how to improve the tracking performance in complex scenes,such as occlusion,fast motion,deformation,etc.is still a problem that needs to be solvedAiming at the problems and challenges of visual object tracking algorithm in complex scenes,we try to design visual object tracking algorithms based on the generative model,discriminative model,hybrid model and deep learning model respectively to obtain a tracking scheme with better universality and performance.The main research contents and achievements are as follows:(1)A generative tracking algorithm based on two level superpixels and feedback is proposed.Firstly,bilateral filter is introduced to filter out outliers and improve the boundary capability as well as segmentation of superpixels.Then,appearance model is constructed based on the coarse level and fine level superpixels.It can adjust the number of superpixels of object and improve the expression power.Secondly,a novel measure method which considers color similarity and relative positions of superpixels is proposed to calculate the confidence map.Finally,through reverse tracking,visual object tracking is designed into a closed-loop system which monitors the tracking process by feedback and update the appearance model.(2)A discriminative tracking algorithm based on multi-scale superpixel correlation filters is proposed.The tracking procedure is treated as optimizing the combination of components of an object.Firstly,a multi-scale superpixel segmentation method is proposed to segment object by the difference of global confidence mask.Then,the confidence of each patch is obtained by calculating the correlation between the candidate of object patch provided by Gaussian distribution based motion model and the corresponding template.A color-feature guided method is proposed for filtering the candidates.Besides,a novel min-max criterion as well as Gene Expression Programming(GEP)algorithm is used to search the optimal combination and estimate the sate of object.Finally,monitoring update strategy is proposed to monitor tracking operation and adjust parameters dynamically.(3)A hybrid tracking algorithm based on the fusion of multi-trackers and multi-features is proposed.A variant Bayesian filter is proposed to combine the feature-level methods with decision-level fusion methods.Firstly,the appearance model of object is constructed by multi-feature,while the motion model is constructed by multi-trackers.Then,Weighted voting strategy and PageRank based strategy are proposed and run parallel to screen the candidates and estimate the state of object.These two candidate selection strategies consider both the inter similarities between template and candidates and the intra similarities among candidates.Finally,a novel decision and update strategy with tracklet prediction-comparison is proposed to cope with the consistency problem of the two candidate selection strategies.The update model can update the appearance model and the multi-trackers in the tracker pool at the same time.(4)A deep tracking algorithm based on attention shake and auxiliary relocation branch is proposed.Firstly,the attention shake layer is constructed by integrating two different attention module by shake-shake network architecture.Then,the proposed attention shake layer is used to replace the pooling layer in Siamese network to improve the expression power.Besides,the auxiliary relocation branch is proposed to assist in object relocation when tracker runs under an untrusted state.According to the prior assumptions of visual object tracking,the structure similarity weight,motion similarity weight,motion smoothness weight and object saliency weight are involved in the auxiliary relocation branch.Finally,based on the response map of the attention shake Siamese network,a switch function is designed to monitor the state and reliability of the proposed algorithm during tracking.