Generic Object Recognition Research Based On Feature Fusion Of 2D And 3D SIFT Descriptors

With the rapid development of computer vision, artificial intelligence and pattern classification, Automatic Object Recognition (AOR) system has become a significant research field of artificial intelligence. Generic Object Recognition (GOR), as one of the major component of the AOR system, has broad prospect of applications in many aspects such as intelligent monitoring, remote sensing, robotics and medical image processing, etc. Due to the large intra-class variations, high inter-class similarities and huge variations of the same object in different views or other reasons in the real environment, the correct recognition rate of the GOR algorithm is usually low. The general features of objects which can express the intra-class commonness and the inter-class differences as much as possible need to be found out. Only extracting stable and effective features can we get the best recognition results in limited training sample. Aiming at these problems, a further research on generic object recopition process is conducted by utilizing feature fusion of 2D and 3D SIFT descriptors. The main work of this paper is described as follows.Feature extraction and representation is crucial to the GOR algorithm,3D SIFT (Scale Invariant Feature Transform) applied in point cloud mode/is proposed in this paper on the basis of 2D SIFT. And the generic object recognition algorithm is completed utilizing multiple feature fusion strategies to combine 2D and 3D SIFT. The specialty of the recognition algorithm proposed in this paper is:Firstly,2D SIFT descriptors of images and 3D SIFT descriptors of point clouds are extracted as the objects features, and BoW (Bag of Words) model is applied to the 2D and 3D SIFT descriptors. Secondly, the SVM (Support Vector Machine) is used as a classification tool. Thirdly, feature fusion is completed by four feature fusion strategies:feature-level fusion and average weighted fusion rule, PCR6 fusion rule of DSmT, Murphy’s rule of combination in decision-Jevel fusion. The final recognition results are given based on the fusion results, thus completing the general object recognition task.Finally, according to a series of simulation experiment results on generic objects, it turns out that the recopition algorithm in this paper can maintain good robustness and high correct recognition rate in the situation of the objects with great intra-class variations, high inter-class similarities or large changes of views, etc.