Indoor Object Recognition Technology Research Based On 3D Point Cloud Data

With the development of artificial intelligence,the application of machine vision has expanded from the traditional manufacturing industry to the indoor service industry.As the core link of indoor scene perception and understanding,indoor object recognition technology has become a research hotspot.At the same time,due to the lower price of 3D scanning equipments and the rise of point cloud processing technology,image recognition has gradually shifted from 2D image to 3D point cloud data,which has 3D spatial geometric features and can naturally decouple objects and backgrounds.Therefore,objects can be recognized better by using of 3D point cloud data in the indoor scenes.How to effectively preprocess the raw indoor 3D point cloud data,select and extract appropriate 3D features to segment scenes and recognize objects accurately,and how to ensure the recognition speed while improving the recognition accuracy have become the focus and difficulties of the research.This paper focuses on the research of indoor 3D point cloud data preprocessing,segmentation and recognition technology.The work of this paper mainly includes the following aspects:(1)According to the current research status,existing problems and development prospect of 3D point cloud scene objects recognition technology,the research scheme plan is proposed.An overall structure of indoor object recognition based on 3D point cloud data is designed,then the 3D point cloud data classical algorithms of preprocessing,segmentation and recognition were introduced.(2)In terms of indoor 3D point cloud preprocessing,aiming at the problem of outliers,undulating noises points and rough boundaries in the raw indoor 3D point cloud data,a hybrid filtering method is proposed to denoise them firstly,to guarantee subsequent features estimate accurately.Secondly,the voxel grid sampling algorithm is used for reducing the number and density of indoor 3D point cloud data,so as to improve the subsequent segmentation and recognition speed.(3)On the part of 3D point cloud indoor sences segmentation,the foregrounds are extracted firstly by RANSAC algorithm firstly.Secondly,two methods are proposed by this paper to solve the adhesive objects segmentation in the indoor foreground.The first is to improve seed extraction,add color information and optimize clustering result on the basis of traditional point cloud region growth algorithm,an improved region growth color 3D point cloud segmentation algorithm is proposed,compared with the traditional algorithm,the stability and accuracy of segmentation are improved.Based on the algorithm of supervoxels concave and convex features segmentation,the second method is to improve the selection rule of seed voxels,the concave and convex features,continuous feature and color feature are fused,a multi-feature fusion of supervoxels 3D point cloud segmentation method is proposed,adhesive objects can be more accurately and completely segmented,providing a good guarantee for subsequent features extraction and recognition of 3D point cloud indoor object.(4)In the aspect of indoor 3D point cloud object features extraction and recognition,the SHOT descriptor is selected as the extracted 3D feature.On the basis of the Hough voting 3D point cloud recognition,the key points are extracted evenly,and the selection of reference points and the rule of feature matching are improved.An improved Hough voting 3D point cloud recognition algorithm is proposed for indoor object recognition.The experimental results show that the improved algorithm can effectively identify indoor objects,and the recognition speed and accuracy are improved on the basis of the original algorithm.This paper contains 39 charts,7 tables and 78 references.