Research On Three Dimensional Object Recognition

2D plane images can be acquired by ordinary CCD camera with very low cost. The research of appearance-based or view-based three dimensional object recognition (3DOR) has been widely researched recently. The view-based recognition methods recognize objects using visual similarity without calculating 3D model for objects explicitly, and that makes the design of recognition system relative simple. On the other hand, the methods based on local feature matching consider the local similarity of views and need not to match all features of views. The local features may be detected on the unoccluded object parts and then the recognition may be accomplished. These methods show good robustness in the condition of object occlusion and background noise. The 3DOR methods based on 2D views and local feature matching, and their applications in the intelligent video surveillance for improving object tracking performance of Kalman filter are investigated. Concretely, the author's research and contributions are listed as follows:1) Multiple kinds of invariant feature extraction algorithms are investigated. The features have mathematical or physical invariant properties including Hu moment invariants, affine moment invariants, color moments, wavelet moments and texture features. Because the continuous wavelet moment can not be used in real application, a discrete wavelet moment algorithm is presented. The simulation results showed that when 3D objects were rotated in views, the resulted moment values of the discrete wavelet moment achieved good rotation invariant property.2) A kind of view-based 3DOR algorithm is researched and presented, using multiple invariant features extracted from 2D views of objects. The features are sent to Support Vector Machines (SVM) for training and recognizing.3) Because a large number of features are extracted, a method is presented which uses Genetic Algorithm (GA) to select feature subset and learn machine optimization at the same time in this 3DOR procedure. One kind of performance boundary of SVM is chose as the fitness function of GA. The boundary can be got once after SVM training and need not extra computing. The evolution time of population thus can be reduced.4) The framework of local feature matching based method is studied. And two descriptor design algorithms are presented. The first one is the improved Affine Invariant Fourier Descriptors (AIFD) based on Maximally Stable Extremal Regions (MSER). Because the AIFD is directly computed based on MSER, the presented algorithm reduces the effects introduced by the image discretization and resampling of image patches during the period of LAF construction. Second, because the descriptors based DCT has the disadvantage of discrimination, the color, texture and shape characteristics of local image patches should be considered. The color moments, texture features and multiple invariant moments can be used to calculate the statistical characteristics of local area and in this way the multiple features descriptor is designed to represent the local appearance of objects.5) To evaluate the performance of various algorithms studied in this paper, a full 3DOR system was developed. The algorithm flow and the effect of parameters selection to the performance were investigated. The recognition performances of the presented methods were evaluated based on two public 3D object image databases. With author's knowledge, the correct rates of recognition based on COIL-100 and ALOI are outperformed to the published results.