Research On 3D Object Recognition Technology In Indoor Environment

At present,the three-dimensional point cloud object recognition methods are mainly divided into two categories: the global feature-based identification method and the local feature-based identification method.The recognition method based on local features mainly relies on key points combined with feature descriptors.By describing the local geometric features around each key point,the features of the target are described.This method has strong robustness to noise and occlusion,but with The increase in the number of objects to be identified will take a long time in practical applications.The global feature-based recognition method relies on the effective segmentation of the scene cloud and the extraction of the target object.This method has strong real-time performance but is sensitive to noise and occlusion.In addition,surface normals are the basic features of point cloud data,which have important influence on the segmentation and feature extraction of point cloud data.In response to the above issues,the main work of this article includes the following aspects:(1)For the errors in the normal calculation process of the abnormal points and noise points in the point cloud data,the point cloud normal line calculation method based on principal component analysis plane fitting is improved.By continuously iteratively calculating the plane equation parameters of the fitting plane and adjusting the threshold value,the outlier and noise points in the point cloud data are eliminated,and the normal accuracy of the high curvature point and the edge point is effectively improved.(2)To extract the clustering surface of the target object from the cloud of the scene more accurately,the traditional European cluster segmentation method is improved.The normal point information and color information of the surface point are added as a constraint to the European clustering segmentation algorithm.The thresholds of the normal,the color threshold and the distance threshold are used to segment the site cloud,effectively avoiding the cloud segmentation of the site.Insufficient segmentation and the disadvantages of over-segmentation.(3)Improved local feature description method Fast point feature histogram,usingobject point cloud data as the object,using(?,?,?)three angle components to represent the normal angle between surface center point and surface point,and According to the angle difference between the normal direction between any point on the surface and its neighboring point,the weight of the point in the surface feature is determined.The three angle components are calculated using the histogram to describe the global features of the surface.(4)An object recognition method combining global features and local features is designed,and the recognition process is divided into two steps: rough recognition based on the global feature-based recognition method,and obtaining the candidate matching result of the target object;and performing on the basis of the candidate result.Based on the local feature recognition process,the target object's precise recognition result is obtained by matching the local features between the object's surface clustering point cloud and the candidate's corresponding three-dimensional model.