The Application Of Deep Learning In Object Tracking

Visual object tracking has been a hot research topic in computer vision underlying many advanced applications in this field.Undergoing complicated occasions in tracking object and background as well as its internal requirement of real time,object tracking has always been a challenging task.Compared with generative appearance model,the appearance mode in discriminatively learned correlation filter has the better ability of feature encoding.Besides,discriminatively learned correlation filters utilizes the efficiency of frequency domain thus this kind of approach has an astonishing tracking speed.As deep learning method have significantly advanced the state of art in object detection,exploiting the deep features in correlation filter framework becomes a new hot research topic.We choose the Spatially Regularized Discriminative Correlation filter(SRDCF)as a baseline tracker and exploit features extracted from deep convolutional neural network trained on object recognition datasets.The outputs of deep convolutional layers encode the semantic information of the targets while the shallow convolutional layers generate high spatial resolution feature map.Therefore,we use the hierarchy convolutional features to locate the target objects.Firstly,we utilize Alternative Direction Method of Multipliers(ADMM)to replace the Gauss-Seidel iterative method for finding the optimal solution of correlation filter,improving the tracking speed and the expansibility of program.Besides,we adopt the Principle Component Analysis to reduce the feature dimension,getting rid of the feature redundancy and improving the tracking speed further.Lastly,we develop a feature fusion method based on information entropy to determine the weights of different response maps.The proposed method shows better results than Deep SRDCF but still has a distinct gap with other related advanced methods.Tracking Methods based on Siamese network have achieved progressive results in arbitrary object tracking.The Correlation Filter Networks(CFNet)enables small scale networks to achieve state of art performance while running at high frame rates via training end to end.However,CF lacks the online training process which results in the insufficiency of use of current video samples.To make better use of current samples to help locate the target objects,we employ generative Gauss Mixture Models(GMM)to produce sample sets facilitating eliminating the redundancy and increasing the variety of samples.The experiment shows that the precision of CFNet2 has been increased by 2.11%and the FPS has been promoted by 13.14%by our proposed method.Besides,the proposed method has a comparative advantage in being associated with on low resolution occasions while maintaining the high tracking speed of CFNet2,but it still faces challenges in terms of the overall tracking performance.