Construction And Application Of Three-Dimensional Geometric Invariant Features From Images

Image based object recognition and tracking are important research fields in computer vision, which are widely applicable in industry or daily lives. Extracted from images, the three dimensional invariants are characteristic representations of the object of interest and invariant to variations in viewpoints, position, illumination, transformations and so on. High level applications such as object recognition and tracking can further benefit from 3D invariants as well. Studies on three-dimensional projective invariants show that there are limited geometric features which 3D projective invariants can be constructed from, such as points. In recent works, the constructed invariants tend to be mainly used in 3D reconstruction and many practical problems remain unsolved. In the context, further researches have carried out in this thesis to the construction and calculation of 3D projective invariants.The main works are summed in follows:Firstly, Methodologies to construct and calculate 3D projective invariants are proposed. Based on the inhomogeneous solution from image point invariants, a homogeneous solution is deduced, which provides a more stable and reliable solution. Moreover, the line features are introduced into the construction of 3D invariants, by which the stability and reliability of constructed 3D invariants are further improved. In addition, extraction of 3D features is made possible even when 3D construction taking place with insufficient information, which profoundly contributes the following applications such as recognition and tracking. In this thesis, the method to construct 3D invariants from point and line invariants is given. Specifically, image features and the establishment of independent equations are described in detail. The proposed method provides a more flexible way to construct and calculate 3D invariants from specially designed geometric structures, instead of the structures introduced in the references.Secondly, image retrieval based on foreground prediction has been proposed. Principles for foreground prediction are established based well studied human attention mechanism and neuron psychology as well as previous image saliency research in computer vision. According to these principles, weights are assigned to image regions to form the most salient foreground. The foreground prediction can therefore constrain the calculation of 3D invariants and provide reliable image information as well.Thirdly, an improved affinity matrix is formed in spectral matching. In this thesis, corners which well defined geometric meaning are used to construct 3D invariants, but corners themselves are lack of discriminate descriptions, therefore the matching can only make use of the geometric relations between corners. We, on the other hand, take full advantage of the color, texture information in their neighboring areas besides geometry and further constrain the matching procedure by forming an informative affinity matrix. A better matching is presented.