| With the continuous maturity of the Internet technology and the rapid development of artificial intelligence,intelligent and automated robots are heavily used in modern production.It is quite necessary to introduce computer vision systems to ensure that robots can accomplish tasks intelligently.As for a robot manipulation system,the first thing to do is to complete the object recognition accurately and estimate its pose from the scene.Compared with 2D image recognition,recognition based on 3D point cloud is a popular research topic in machine vision because of additional depth information.This thesis analyzes the current development trend of 3D object recognition,focusing on the key technologies of 3D object recognition and pose estimation based on Kinect.The scene cloud is collected using the Kinect sensor,and then target identification and the pose estimation will be finished through the method of local feature matching.It can be divided into four algorithm modules: cloud acquisition of field sites,3D reconstruction of models,3D target recognition,and attitude estimation.The specific tasks include:1.3D reconstruction of field sites:To solve the problem of the noise and voids existing in the depth image,there are two steps to be done.First,the scene depth image and the color image are obtained with the Kinect sensor.Second,the bilateral filter algorithm is used to remove the holes and noise on the depth image.After that,a better quality scene 3D point cloud is acquired with the depth image and the color image being registered.2.3D reconstruction of models : Firstly,an omni-directional three-dimensional reconstruction of the scene where the target object resides is accomplisehd by using the Kinect Fusion scene reconstruction algorithm.Then,the target object is extracted by the point cloud post-processing to realize the omni-directional 3D reconstruction of the model and establish a model library.3.3D target recognition: In order to achieve accurate target recognition,the first thing to do is that the scene spot cloud and the model point cloud should be sampled uniformly and the key points are extracted.Meanwhile,the SHOT local feature descriptors of the key points are established.Second,the Kd-tree is used to search matching point pairs as well as the 3D Hough voting algorithm is introduced to find the pseudo correspondence pairs.4.Attitude estimation:The ICP algorithm is used to calculate the rigid transformation matrix of the identified target,and the target attitude parameters are finally obtained.At last,an experimental platform was built,and the proposed algorithm was tested and verified.The experimental results proved the reliability of the algorithm. |
