Research And Implementation Of Object Recognition Based On RGB-D

Object recognition is a basic and important research direction in the field of computer vision and pattern recognition.Object recognition technology plays an extremely important role in many computer vision research and applications,such as service robot target recognition system,intelligent monitoring equipment and intelligent traffic identification system,etc.,which is a hot issue in various fields now.In recent years,object recognition based on RGB images has already developed a mature study.However,due to the lack of spatial information of objects,the object recognition effect is not very ideal.As Microsoft has launched a new generation of cameras,Kinect,two dimensional color images(RGB)and distance information(Depth)can be obtained at the same time,researchers can further study object recognition through three-dimensional information(RGB-D)of objects.Based on the RGB-D information obtained by Kinect,this paper studies the technology of object recognition.Aiming at the problem that the feature fusion in RGB-D object recognition research fails to make full use of RGB-D's advantage and the low precision of object recognition,some improved methods and measures are proposed.The main research contents of this paper are as follows:(1)Since the existing artificial feature extraction methods have the problems of weak generality and high cost of feature construction,this paper uses an improved hierarchical feature learning method to extract RGB features and depth features respectively.Firstly,the KSVD dictionary learning algorithm,sparse coding and spatial maximum pooling are used to extract the underlying features of the image.Then the above methods are used to process the underlying features to obtain the high-level features of the image.(2)Most of the existing feature fusion methods are simply to combine RGB features and depth features to obtain comprehensive features and then classify and identify them.There is no difference between the two modal information,and the advantages of RGB-D cannot be fully utilized.In this paper,a multi-modal feature fusion algorithm based on SVM is proposed.According to the contribution of different modal features to the final recognition result,different weight coefficients are assigned to them,and a larger weight is assigned to the modal features with larger contributions,the small modal features are assigned smaller weights,and then fused to obtain more expressive synthetic features,which makes full use of the advantages of RGB-D.It is also proposed to use the simulated annealing algorithm and the least squares fitting method to optimize and solve the weight parameter and the final classification recognition rate.(3)By combining the hierarchical feature learning method and the multi-modal feature fusion algorithm based on SVM,an RGB-D object recognition method based on hierarchical feature learning and feature level fusion is proposed.By experimenting on a standard RGB-D dataset,the results show that the proposed method can obtain higher accuracy of object classification and recognition than other existing methods,and the validity of this method is verified.