Research On 3D Object Recognition Based On Point Cloud

Object recognition is an important research direction in the field of computer vision.In the past few decades,2D image recognition has been extensively studied and has achieved fruitful results.Compared with 2D images,the 3D point clouds are not affected by illumination intensity,rotation transformation,and the scale changes,and can provide richer geometric information,therefore,3D point cloud recognition can estimate the pose of the object more accurately.In addition,with the rapid development of data acquisition equipment,it is easier to obtain point cloud data.All these advantages make 3D point cloud recognition a research hotspot.It has been widely used in many professional fields.The paper makes a lot of research on previous work,for the recognition accuracy,efficiency and robustness to noise,two effective 3D point cloud recognition algorithms are designed.Firstly,the paper introduces the background,significance,application prospects and research status at home and abroad of 3D point cloud recognition.Meantime,the paper introduces the acquisition method of point cloud,the preprocessing of point cloud and the related theoretical basic knowledge in the process of point cloud recognition in detail,which lays the foundation for the subsequent recognition work.Secondly,after analyzing the Hough transform in 2D image,this paper extends it to three-dimensional space and puts forward a 3D object recognition algorithm based on Hough voting.Firstly,the original data is reduced by uniform down-sample,and then the key point extraction algorithm based on the normal vector projection mean is designed.The key points are described in combination with the SHOT feature,and the initial correspondence between the point pairs is determined according to the geometric relationship between the descriptors.The three-dimensional Hough space voting strategy is used to eliminate the wrong matching relationship and complete the object recognition,then use a cubic box to select the target object.Finally,in order to facilitate human-computer interaction,a human-computer interface for 3D object recognition is designed based on Qt.Then,based on the predecessors' work,a point pair feature recognition algorithm based on key points is proposed.Firstly,the original point cloud data is down-sampled by using the voxel grid method,and the key points are extracted by the ISS algorithm.Then use the set of key points to build point pair features,quantized and stored in a hash table to build a global descriptor.In the 2D space,the model point cloud and scenepoint cloud are matched and identifiedby using a fast voting strategy,and finally the object pose is optimized by ICP.The algorithm combines a global model descriptor based on point pair features and a local matching method based on a fast voting strategy,and without segmenting the scene.The extraction of key points reduces the search range of point pair features,and improves recognition efficiency and effectively suppresses noise interference.Finally,in the VC ++ environment,the experiments are carried out on different algorithms proposed in multiple datasets and real scenes to verify the feasibility and effectiveness of various methods proposed in this paper,and the robustness to noise is strong.The algorithms proposed in this paper have some practical engineering application value.