Research On Feature Extraction And Tracking Technology For Extended Objects

In the field of computer vision, extended objects recognition and tracking is an active research point. Feature extraction is an important part in the process of object recognition and tracking, which directly influences the performance of later recognition and tracking. Therefore, it is valuable for reality life to research feature extraction and tracking technology for extended objects. However, since extended objects have some characteristics that are different from generic objects, general object tracking algorithms cannot be used for extended objects. Moreover, the existing factors during tracking, such as 3D pose variation of object, image blur, illumination change, rotation, occlusion and partial deformation, are huge challenges for feature extraction and tracking of extended objects. In this case, the task is proposed to solve these problems.In this paper, the extended objects are obtained via photoelectric detecting system. On the one hand, for extended objects with small pose variation, blurring edge and unapparent texture at long distance, this paper studies the algorithm of high precision localization; on the other hand, for extended objects with big size, texture and shape at close distance, the paper exploits invariant feature extraction and robust tracking algorithms. The main contents are as follows:(1) For extended objects at long distance, this paper studies the algorithm of high precision localization in depth. On the basis of the geometric shape feature of objects, a high precision localization algorithm for profiled extended object based on General Hough Transform(GHT) is proposed. Numeral experiments on simulant and outfield sequences have verified the effectiveness and stability. The localization precision is up to sub pixel, which is less than 0.6 pixels.(2) For extended objects at close distance, invariant feature extraction and pose computation technology are exploited deeply. Firstly, current state-of-the-art feature extraction, detection and description algorithms in home and board are analyzed in detail. Moreover, numeral experiments are performed to compare the performance of different algorithms. Secondly, analyzing the characteristics of the extended objects at close distance, we use skeleton to describe the geometric structure and topology of the object and propose a context descriptor based on skeleton and contour. Experiments show that the proposed descriptor can achieve stable matching during pose variation. Compared with shape context descriptor, the matching speed is largely improved.What’s more, for later object pose computation, given actual size of the object and the parameters of camera, a 3D pose computation algorithm is developed on the basis of skeleton and geometric projection relation in the view of mono-vision. Some experiments validates the effectiveness and stablility of the proposed method.(3) Aiming at automatic recognition of the critical part of plane at close distance, multi-rectangle feature is exploited based on Haar-like in depth, and we use Adaboost classifier to learn the best features such that the critical part of the plane can be recognized automatically, intelligentizing tracking.(4) For extended objects at close distance, the influences of feature selection and integration on tracking are investigated in depth. Moreover, to obtatin stable tracking of extended objects, we combine skeleton and distribution fields. At the same time, for the sake of adapting to rotation and scale variations, we improve the distribution fields in two-folds: first, aiming at the problem of fixed layers in original algorithm, heterogeneous adaptive delamination is developed, which reduces computation complexity and storage while retains good performance. Second, the original algorithm cannot adapt to rotation and scale changes. To solve this problem, we integrate BRISK detector to the algorithm to obtatin robust tracking.(5) For extended objects at close distance, the influence of appearance model on tracking is investigated in depth and extended object tracking is formulated in the view of modeling. The influences of sparsity collaboration model and context model on tracking performance are explained in stress. Then based on context model, we integrate BRISK feature to the model to achieve scale and rotation adaptation.In summary, this paper exploits some critical problems in extended object recognition and tracking such as invariant feature extraction, high precision localization for extended objects, feature integration, appearance modeling, automatic recognition of critical part of the plane, and stable tracking, etc. achieving robust tracking for extended objects with different types.