Research On Object Recognition Method In Cluttered 3D Point Clouds

The object recognition in the 3D point cloud has always been a hot topic in the field of computer image and pattern recognition.However,there are a lot of nuisance in the 3D point cloud,such as noise pollution,clutter and occlusion.The object recognition based on local feature can solve the problem of object occlusion in clutter scene,but the noise and point density variation has been a very challenging problem in object recognition in 3D point cloud.In view of the problem of object recognition in 3D point,this paper mainly completes the follow work:First,An overview of the framework for 3D object recognition based on local features is presented,and each step is given in detail.This paper makes a lot of research on the key steps that affect the efficiency of object recognition: interest points and description extraction.Second,In the step of interest point extraction,we mainly study the interest points method of Intrinsic Shape Signatures(ISS3D),which is very sensitive to noise.In the establishment of Scatter Matrix,we use distance and point density weight to deal with noise and point density variation.And the idea of tensor voting is introduced into the interest points extraction process,and the steps of removing the outlier points are added.The experimental results show that the improved ISS3 D interest points method has a superior effect on noise and point density variation.Third,In the stage of feature description,this paper mainly studies the signature of Histograms of Orien Tation(SHOT).SHOT has inherited high descriptiveness based on signature method,and has high robustness based on histogram method,which is widely used in 3D object recognition.SHOT has a high robustness to noise,but it is very susceptible to the change of point density variation.In this paper,a new geometric center is proposed when establishing local reference frame(LRF).When calculating the SHOT histogram feature,the minimum feature direction corresponding to the improved scatter matrix issued instead of the normal direction.Experiments show that the improved SHOT feature has higher robustness to noise and point density variation.At last,In theobject recognition experiment in clutter3 D point cloud,this paper firstly deduces the three kinds of point cloud densities in the Bo D1 model library with only noise intensity change,and generates the simultaneous existence of noise and point cloud density Of the model library.In this paper,we use the nearest neighbor compared to the second nearest neighbor method(NNDR)to compute the correspondence relation.Then we use the generalized Hough Transform(GHT)to generate the hypothesis and use a global hypothesis verification method tofill out all the assumptions,and choose the best place where the object may appear in the clutter scene.Experiments show that the proposed method has a superior effect on the noise and point cloud density variation problem in the clutter scene.