Research On Key Technology Of Object Tracking Based On Multi-feature Fusion In Complex Scenes

Visual object tracking(VOT) is a hot research topic in the computer vision, image processing and pattern recognition community. It has many promising applications including intelligent surveillance, human computer interaction, intelligent transportation, military guidance and robot navigation, etc. In complex scenes, designing an effective visual tracking system, however, is a challenging task due to object appearance variations caused by illumination changes, background clutters, object deformation, and occlusions, etc. In recent years, although many domestic and foreign scholars have carried out extensive and deep research in visual object tracking field, and present a lot of excellent algorithm, but there are still many key issues unresolved, designing a real-time visual object tracking system with more robust and higher tracking accuracy is an urgent need for current military and civilian sectors.In this thesis, under the framework of the classical particle filter and Mean Shift theory in visual object tracking field, we researched the key technology of object tracking based on multi-feature fusion. To improve the robustness of object tracking algorithm as the main goal, in view of the problem of low tracking precision and large tracking error in complex scenes, the classical object tracking algorithm is improved by the multi-feature fusion strategy. The main innovation research contents and contribution of this thesis are as follows:(1) A construction scheme of target tracking framework based on multi-feature fusion is proposed, the overall framework of target tracking based on multi-feature fusion is established, a generic model of multi-feature fusion target tracking has been designed, by introducing the function of measuring the feature description ability, the multi-feature fusion strategy based on the measuring of the feature description ability is proposed, which is the basis of achieving multi-feature fusion tracking under different algorithm framework.(2) A novel particle filter object tracking method based on adaptive multi-feature fusion is proposed. In order to solve the poor robustness problem of tracking algorithm based on single feature or fixed weight multi-feature object model in the object tracking process, an adaptive fusing multi-feature tracking algorithm is proposed based on the discriminability and stability of features in the particle filter framework. Several reliable features are adaptively selected by calculating their discriminative ability and stability, which are used to describe the object model, establish the multi-features fusion object model and set the importance weights of features. In the process of state transition, a selective template updating method is presented based on the measurement of feature stability, and the occlusion problem is handled. Experimental results show that the proposed method can track object under complex scenes in robust performance.(3) A scale and direction adaptive Mean Shift object tracking method based on multi-feature fusion is proposed. The classical Mean Shift algorithm only uses the color feature to describe the object model, lacking an adaptive update mechanism of the object appearance model, which will reduce the tracking accuracy, In order to solve the problem in the object tracking process, a scale and direction adaptive Mean Shift tracking algorithm based on multi-feature fusion is presented in complex scenes. In the process of the dynamic changes of the scenes, several reliable features are selected according to their ability to distinguish between object and background, by which the object model is represented, the multi-feature model is established and the feature importance weights is set. Mean shift object localization formula based on multi –feature fusion is proposed. By evaluating the reliability of the different features in different scene dynamically, update the feature weight and fuse the multi-feature adaptively. According to the weights of different features, a selective template updating mechanism is put forward to alleviate the model drifts. Mean Shift algorithm usually use a fixed bandwidth of kernel function, which has poor adaptability to the scale and direction change of the object appearance model, so the SIFT(Scale Invariant Feature Transform) feature is introduced into the Mean Shift algorithm framework, which make the algorithm has good adaptability to the scale and direction change of the object appearance model. Experimental results show the robustness and effectiveness of the proposed method in complex scenes.(4) A novel patch-based tracking method based on the local sensitive histogram(LSH) and superpixel model is proposed. In order to address the problem of losing object based on the holistic appearance model in complex scenarios, LSH feature and adaptive patch-based method were used to establish the object model, by extracting the illumination invariant feature of the LSH resist the influence of the illumination changes on the object model effectively; for the lack of effective occlusion handling mechanism of the LSH algorithm, the patch-based method based on the superpixel segmentation is introduced to improve the performance of resistance occlusion; by through the relative entropy and mean shift cluster method, measuring the differences confidence value and the foreground-background confidence value of the local patch, establish the dual weights constraint mechanism, which is used to calculate the confidence value of different patches, the patches with high confidence value are selected to locate object in the particle filter framework. Comprehensive experiments on the benchmark set confirm that the proposed method outperforms the state-of-the-art algorithms in many cases in complex scenarios.