Research On Object Tracking Algorithm Based On Support Vector Machine Via Multi-Feature Representation

Visual object tracking,as a key technology for video content analysis,has been applied to many fields with broad application space and good development prospects,such as video surveillance,human-computer interaction,intelligent navigation and medical diagnosis.In recent years,discriminative tracking algorithm has extensively studied in visual object tracking due to its strong robustness.It constructs the discriminative models by making full use of the object and background information on the images.Based on the deep analyses of discriminative tracking algorithm,this thesis intensively study on the classical support vector machine(SVM)method and the latest support correlation filter(SCF)method.The main research contents of this thesis are as follows.(i)The tracking mechanism of discriminative tracking algorithm is analyzed,and the mathematical models and kernel extensions of SVM and SCF are deduced,respectively.There is a test dataset and some quality evaluation methods of target tracking algorithm in this thesis.Both the analysis and the dataset provide the theoretical basis for method research and experimental analysis.(ii)Since the performance of the traditional SVM-based method is severely limited by its feature representation that single hand-craft feature representation method cannot enough represent target information.Firstly,this thesis analyzes the main appearance information of the target and the characteristics of the different hand-craft features.Considered the complementarity of the histogram of oriented gradient(HOG)feature and the color name(CN)feature for the target's gradient and color information,a feature integration with HOG and CN is proposed to sufficiently represent the target,and a SVM model based on multi-feature integration is constructed by using a linear kernel function.Secondly,in order to avoid the SVM model's incorrect updating due to the occlusion interference,this thesis analyzes the change relationships of the target between before and after occlusion,and formulates an update strategy based on similarity discrimination.Based on the research mentioned above,a novel support vector machine-based tracking algorithm with feature integration is proposed.Finally,the effectiveness of the proposed algorithm and the update strategy are verified by comparative experiments,respectively.(iii)In order to deal with the model drift of the SCF method,which is caused by boundary effect,the thesis first analyses the essential reason of the model drift,starting with the SCF model.Then the human visual attention mechanism is introduced into building the constraint relationship between the object and background on the image,which can strengthen SCF model's attention on the target area and penalize its attention on the background area.By exploiting the priori information between adjacent frames of the continuous video sequence,this thesis introduces a temporal regularization term into the SCF model,and constrain the error of its parameters between adjacent frames.Based on the research mentioned above,a novel object tracking algorithm using spatial-temporal regularized support correlation filters is proposed.Finally,the effectiveness of the proposed method is verified by the experiments in the holistic contrast and the background clutters contrast.