The Research Of 3D Object Recognition Method On Fusing RGB And Depth Features

3D object recognition is one of the hot issues in pattern recognition and machine vision which has been popularly used and manifested with huge economic values in military, transportation, biomedical and may other fields. To solve the existence of limitations which based on RGB information and Depth information in recognition problem and the existing uncertainty problems of intra-class differences and inter-class similaritis between the objects, our research of 3D object recognition method on fusing RGB and Depth features is conducted in the feature layer and decision layer and all the experiments images are obtained by Kinect camera.The main work of this thesis is illustrated as follows:(1) We analyze and summarize the current related work of object recognition using RGB information and Depth information, and the publicly available datasets for 3D object recognition were introduced in detail.(2) In the feature layer, different features importance weights between the object to the result of the recongnition were introduced, then adaptive weighted fusion of RGB and Depth features by using multiple kernel learning meathod, then combined SVM classifier to achieve 3D object recongnition. Several experimental results on the RGB-D datasets demonstrated the effectiveness of the method in fusioning of the features, and largely solved the impact on the recongnition results of issues about the intra-class differences and inter-class similaritis, improved 3D object recognition rate.(3) On the decision level, for the limitations of using single features and the existing uncertainty problems during recongtion process. A 3D object recognition method based on fusion of RGB and Depth feature using D-S evidence theory was researched. Combined the advantage of SVM classifier has good classification ability when the samples number is small, then we done a judgement on the probability output of two types single features of SVM were fused by using D-S combination rule. Experimental results on the RGB-D datasets showed that this method could effectively improve the recognition rate of 3D objects.