Research On Object Detection And Pose Estimation Methods For Robot Opertion

With the development of science and technology,robots are not only used in industry,but also into people's lives.At the same time,higher requirements are placed on robot technology.Unlike the industrial robots of the last century,most of today's robots have a certain degree of intelligence,and can perform various tasks well even in an unstructured environment.Object detection and pose estimation are key technologies for robots to realize autonomous operation,and they have important research value and application prospects.There are many methods available to solve the problems of object detection and pose estimation.This article uses image feature-based methods for target object detection and pose estimation.The main research contents are as follows,(1)A new feature point algorithm,Multi-Neighborhood Structure Tensor Features(MNSTF)algorithm is proposed.The mainstream feature point algorithms mostly only rely on the texture information of the image,but ignore the structural information of the image,so they cannot solve the problem of no texture.The multi-neighborhood structure tensor feature algorithm uses a series of fixed feature point neighborhoods and the image local structure tensor,so that the MNSTF feature descriptor not only expresses the texture information,but also expresses the information of the image structure level,and realizes the MNSTF descriptor.Rotation invariance,high discriminability,versatility and robustness of,solve the problem of weak texture and no texture.(2)A stereo matching algorithm based on MNSTF features is realized.In the field of binocular vision,it is often impossible to accurately calculate the spatial position information of pixels due to the weak texture of the image.The MNSTF algorithm can not only effectively extract the features of rich texture images,but also the features of weak texture images.Therefore,in this article,we try to apply the MNSTF algorithm to the registration of pixels in binocular vision,and combine the constraints of binocular vision with MNSTF.The organic combination of feature algorithms can obtain more accurate location information.(3)Object detection and pose estimation based on MNSTF algorithm.Object detection and pose estimation are the focus of the subject research.The MNSTF feature point algorithm can achieve robust detection of image feature points and correct matching of feature points in the case of weak texture,which can be used to solve this problem.First,the target object is detected by matching with the database feature points;then,the ICP algorithm is used to estimate the relative pose of the object and the database object.On the basis of the above research,the subject has carried out a large number of experiments,including the matching experiment of MNSTF feature descriptor,the comparison experiment with the SIFT algorithm,the occlusion experiment,the stereo matching algorithm based on the MNSTF feature descriptor,the ICP algorithm experiment,and the ZED double The eye camera establishes an experimental system for object detection and pose estimation.In the experiment,the effectiveness of the MNSTF feature point algorithm and the feasibility of the object detection and pose estimation method are verified.