Research On Correlation Filter Based Visual Object Tracking

Visual object tracking is a fundamental research topic in pattern recognition and computer vision,and is a significant processing technique in image and video analysis,with undiminishing practical applications in artificial intelligence.A tracking system aims at establishing dynamical appearance model in complicated scenario,performing predictions to the target state and locating the target continuously in a video sequence given its initial state.Over the decades of research development,numerous algorithms have been proposed for designing tracking theories and methods.The increasingly challenging tracking bench-marks with corresponding evaluation methodologies also help to ensure a sustained vitality of the research area.In recent years,due to the high performance achieved by the mod-ified machine learning methods and deep neural networks in various visual tasks,visual object tracking community also focuses on these two topics.Specifically,discriminative correlation filter based tracking methods employ data augmentation technique supported by time-frequency transform theory with improved tracking efficiency.On the other hand,deep neural networks provide more discriminative feature representations compared with traditional fcature deseriptors,achieving further enhanced precision and robustness.In this dissertation,the author analyse the theoretical deficiency in filter modelling process,explore the combination forms of feature representations,establish novel correlation filter modelling techniques,and propose effective implementations.The main contributions include:(1)A novel discriminative correlation filter model based on adaptive spatial feature selection is proposed.Group sparsity,temporal consistency and thresholding operators are employed to achieve feature selection.The selected spatial features support improved discrimination,alleviating boundary effect in the traditional correlation filter framework.Both foreground and background appearance is considered in filter modelling process,obtaining filters lying in a low-dimensional manifold with enhanced interpretability and robustness.(2)A joint spatial and channel feature selection discriminative correlation filter learn-ing model is proposed.Both spatial units and channel attributes are grouped with low-rank and sparsity constraints to train the discriminative correlation filter.This method performs feature selection in two dimensions,obtains the filters with lower dimension,simultaneously alleviating boundary effect in the spatial domain and decreasing noise as well as redundancy in the channel domain,which further improves the tracking perfor-mance.(3)An accelerated correlation filter tracking model based on improved alternative di-rection multiplier optimisation is proposed.Light-weight deep neural network is employed to realise feature extraction,and continuous dynamical system is fused in designing dis-crete iterative optimisation method.Based on the proved equivalence between discrete iterative optimisation and continuous dynamical system,iterative momentum and adap-tive initialisation are introduced in the traditional alternative direction multiplier method,the algorithm achieves a convergence rate improvement from O(1/k)to O(1/k2),with guaranteed stability provided by the theory of continuous dynamical system.(4)A coarse-to-fine tracking strategy based on the diversity between different features is proposed.Hand-crafted and deep features are explored and analysed in terms of tracking precision and robustness.Deep representations are employed to perform coarse tracking due to its superiority of larger respective field,while hand-crafted feature descriptors are then used to achieve fine grained tracking with larger spatial resolutions.To improve the tracking preformation with current learning models,the diversity of different feature representations is studied to achieve a two-stage complementary tracking strategy.