Object Recognition Algorithm Based On RGB-D Visual Information

Object recognition is an important research direction in the field of machine vision,which involves many disciplines such as mathematics,machine learning,image processing and pattern recognition.With the development of society and the maturity of machine vision technology,the application of machine vision in the field of service robot has been expanding,which has promoted the process of national industrial Strategy 2025.Object recognition algorithm based on rgb-d information is an important part of service robot system,and it is widely used in intelligent driving,UAV and humanoid robot.Therefore,the research of object recognition in the field of machine vision is of great theoretical significance and practical value.Object recognition based on rgb-d(red: Crimson Channel,Green: Blue channel and depth: Deep channel)information is an important research topic in machine vision field.The object recognition algorithm includes two parts: feature extraction and classifier design,and how to design a better feature extraction algorithm has always been a hot topic in people's research.In this paper,two kinds of feature extraction algorithms are proposed,one is the object recognition algorithm based on multi-channel dictionary,the other is the object recognition algorithm combining sift and sparse coding,the main contents include:(1)Multi-Channel feature dictionaries for RGB-D object recognition is proposed.In the field of object recognition based on RGB-D information,traditional feature learning algorithms usually study RGB three-channel color characteristics as a whole.When the RGB three-channel color information is extracted by sparse coding algorithm,single dictionary is obtained.Compared with the single dictionary,the multi-channel dictionaries can extract richer image block feature information and can express the feature of image block more accurately.Each image is divided into several blocks,each divided into several units,each containing several pixel points.When extracting features,we first use the dictionary to solve the sparse coding of each pixel in the image,and then use the maximum pooling algorithm to get the cell features.A block feature is a concatenation to all of the cell features it contains.The block feature is the first layer feature,and the second layer dictionary and its sparse encoding can be obtained based on the block feature.The first layer feature representation and the second layer feature representation can be obtained by using spatial pyramid maximum pooling respectively for block feature and its corresponding sparse coding.The experimental results show that the object recognition algorithm based on multi-channel dictionaries are more accurate than that of object recognition in single dictionary,and the second layer feature is more accurate than the first layer feature.(2)Combining SIFT and sparse coding for object recognition is proposed.Many sparse coding algorithms are mainly based on the color,spatial and shape information of image blocks,while the gradient direction information is neglected.The SIFT feature can extract the gradient direction histogram statistic information of the image block and has the scale rotation invariance.The SIFT feature extracted from the image block can increase the image block gradient direction information based on sparse coding.Each image is divided into several overlapping blocks,then extracts the SIFT feature from the grayscale image block and the corresponding dictionary and sparse coding are obtained.Using sparse coding,the block feature and sparse coding based on block feature are obtained directly from color and depth images.The pyramid pooling algorithm is applied to get three kinds of image features on sparse coding based on sift,block feature and sparse coding based on block feature.Finally,the three kinds of image features are linked to the final expression of the object features.The results show that the method is effective.The results show that our method is effective.