Research On Adaptive Visual Target Tracking Method Based On Correlation Filter

As a hotspot in the field of Computer Vision,visual tracking has been widely used into military and civil problems,including intelligent driving,behavior analysis,medical diagnosis and some other practical applications.Visual tracking aims to enable computer to simulate the human visual system and the way of thinking,in order to identify the moving objects in the field of vision,and finally achieve an accurate location.Although great progress had been made in Computer Vision,the real situation is usually complex and changeable,so not only accuracy but also tracking speed is required for the tracking algorithm.There is still a long way to go on applying the academic tracker into practical applications.The correlation filtering based tracking methods(CFT)has gradually occupied a dominant position with the advantages of both accuracy and speed among the existing tracking algorithms.In order to solve the tracking problems in terms of real-time requirement,complex tracking conditions and changeable motion state,three novel technologies werer proposed in this paper from the three perspectives,including robust appearance model construction,improvement of target location,and dynamic modeling for the change of foreground and background.The main research achievements include:(1)Target appearance representation based on evolutionary feature subsetThis method proposed to compare the whole feature set as a population and the single feature vector as an individual in the evolutionary algorithm,then construct an optimization framework based on evolutionary algorithm to filter the optimal subset of multi-feature.The target appearance model trained by eliminating redundant information could be more robust.The idea of this feature optimization method is that,if two or at least two kinds of feature are utilized for tracking,then the optimal feature combination for tracking can be analyzed and optimized according to the characteristic of the current video sequence.The optimization method enriches the diversity of species by the operations such as selection,crossover and variation operation between individuals,and leave the better individuals according to the objective functions in the processing of iteration.As a result,the dimensions of feature vectors are decreased by filter the redundant information against for tracking,and finally the robust target appearance model can be trained with the optimal feature combination.Compared to the correlation filter based tracker without filter redundant information,the method proposed in this paper is more robust with less computing burden,not only improved the tracking accuracy but also enhanced the tracking speed.(2)Adaptive energy aware localization for visual trackingFor the correlation filter based trackers enable to accurate detect the insignificant target in the Furious domain,this paper proposes to automatically detect the imbalance between the foreground and background in the process of localization,and adjust the energy distribution adaptively.In the method proposed,saliency detection is used to extract the accurate contour of the real target for generating the image mask,which could separate the foreground and background precisely.This method analysises the distribution information between foreground and background in realtime.If the energy of frame to be tracked is imbalanced,adaptive adjust its energy distribution and then provide the sample to the correlation filter for model training.Do correlation between the correlation template trained by all the tracking results and the detection region in current frame;localize the target according to the response map obtained.For the training of correlation filter,our paper also proposes a topdown and bottom-up combined training strategy,reject the wrong samples in time to ensure the reliability of the training set,finally train and update the correlation filter according to the weights of samples.The proposed tracker achieves a great improvement on tracking accuracy than the traditional correlation filter based tracker;the robustness has also been significantly improved in dealing with various tracking difficulties.(3)Adaptive dynamic modeling method for target stateThe appearance of non-rigid object is quite complex and changeable in the process of tracking,so how to model the changeable appearance of the target is still a challenging problem.In this paper,a tree structure is utilized to maintain the correlation filter templates in video sequence,in which the representative states are stored in nodes,the relationships between templates are represented as edges in tree.A stable tracking is realized by contentiously constructing and updating the tree structure in the tracking process.Our method treats the model update as a match of different target states;every representative state is trained to a correlation filter template separately.A reliability analyze for different target states is achieved by evaluating the qualities of the correlation filters,finally a dynamic model is obtained according to the most reliability information.Our algorithm records the target states in the tree structure,and stores all the correlation filters trained by every stable states in nodes.By analyse the stability,reliability and different relationship between states,our method assigns every node and edge a weight.The construction and optimization of the tree structure are realized according to the weights of nodes and edges.Finally,the proposed method can effectively avoid tracking with unstable information,and uses the most reliable information for tracking.By testing this correlation filter tracker based on dynamic modeling,it can be concluded that this method could achieve a higher accuracy and be more robust than the traditional correlation filter based tracker.