Visual Tracking Based On Feature Learning And Feature Fusion

Visual object tracking is a fundamental research field in computer vision.It plays an important role in various nowdays artificial intelligent applications,such as smart video surveillance,automated driving cars,unmanned aerial vehicle navigation,and home robotic.As one of the key issues in computer vision,its breakthroughs are of great signifi-cance in promoting the development of a new generation of artificial intelligence industry in China.The representation of the tracking target is a basis of all visual tracking methods:generative methods need it to build target models,discriminative methods need it to learn classifiers,and CNN-based methods exploit the powerful feature learning ability of CNNs.Therefore,a powerful representation of the target facilitates the tracking meth-ods to achieve top performance.We conduct researches on feature learning and feature fusion in the following four aspects:1.We propose a structure-aware local sparse coding model to address the loss of spa-tial structure in the encoding process of traditional local sparse coding model.The pro-posed model not only preserves the local sparsity but also requires that candidate patches should be represented by patches only from templates that are similar to this candidate,which can thus preserve the spatial structure among selected template patches.To corpo-rate this constraint into local sparse coding model,we simultaneously encode all patches of a candidate,and design a rearrangement function to rearrange coefficients correspond-ing to the same template into a same row.Then,we can preserve the spatial structure by minimization the _2,1of the rearranged coefficient matrix.We apply this model to vi-sual tracking tasks,and extensive experimental results demonstrate the effectiveness of the proposed method.2.We propose an attribute-based multi-branch convolutional neural netwok?CNN?to address the feature learning problem of CNNs when only a small amount of training data is available.The proposed model first exploits an individual CNN branch to learn representations from each attribute group,and then utilizes ensemble layers to learn how to adaptively combine these representations together.This design facilitates the feature learning compared to existing methods that utilize the same CNN layers to learn represen-tations of objects under different challenging attributes.Additionally,the data grouping operation reduces the appearance diversity fed into each CNN branch,and hence also makes the feature learning easier.Extensive experimental results demonstrate the effec-tiveness of the proposed method compared to state-of-the-art tracking algorithms.3.We propose an adaptive cumulative regret model for the Hedge algorithm to take full advantages of different CNN features.The cumulative-regret-based Hedge transfers the decision loss of each expert into a measurement call'Regret'and updates decision weights of experts by minimizing the cumulative regret of each expert.We maintain an adaptive cumulative regret model for each expert,which adaptively adjust the ratio of historical regret to the current regret according to the performance changes of each expert.We evaluate the proposed ensemble tracker on large object tracking datasets and excellent performance are achieved.4.We propose an adaptive instantaneous regret model for the Hedge-based ensem-ble tracker,which not only combines the advantages of multiple features but also has an upper bound of the error.We improve the Hedge algorithm from three aspects.First,we propose an preciser loss estimation of each weak tracker by considering both the appear-ance similarity and the spatial distance,where we design a Siamese network to compute the appearance similarity between tracked results of each weak tracker and the target tem-plate.Second,we propose a new potential function based on one-order cumulative regret,which generates much smoother weight distributions on weak trackers.Last,we change the adaptive cumulative regret to adaptive instantaneous regret,which bounds the changes of cumulative regret in a controlable range,and we prove that this design indeed leads to an upper bound of the loss of the Hedged tracker.Extensive results on large benchmark datasets demonstrate the effectiveness of the proposed method compared to state-of-the-art trackers.