Research On Door Handle Recognition And Pose Estimation Based On Vision

As an important research direction,machine vision is getting more and more attention.In an unstructured environment,the robot only operates after the target is cognized.The purpose of this paper is to solve the problem of the identification and attitude estimation of the door handle in the deslagging operation of the metal magnesium reduction tank.The accuracy of the identification and clamping position of the door handle directly affects the success of the robot door opening.Changes in the external factors,such as the position,angle,spacing,and lighting of the target,can lead to reduced system identification accuracy and increased pose estimation errors.Therefore,it is necessary to study the accuracy of the door handle recognition and space attitude estimation.This article studies the identification of the door handle based on the binocular vision system.It is aimed at the problem that the position and attitude of the door handle in the industrial site are varied.This paper uses the RBF neural network algorithm to identify the image with good fault tolerance.The problem that the accuracy of the door handle recognition will be reduced due to the complexity of the door handle background,the exposure and other external factors.In this paper,the RBF neural network combined with 4-neighborhood segmentation method is proposed to realize the segmentation of the interference region in the image under the condition that the expected effect cannot be achieved in the early stage,and then to eliminate the other interference objects and accurately identify the door handle.All the problems encountered in the industrial field can be solved well and the reliability of the application of this algorithm is verified.There are two aspects to the study of attitude estimation methods based on vision-based door handles: position and attitude.Therefore,the position of the door handle needs to be determined first,and the space gesture of the door handle is determined based on the relationship between the coordinate systems.The door handle is already fixedly installed,so the roll angle is zero.In order to eliminate lens distortion and errors and determine the camera's internal and external parameters,camera calibration is required.In order to eliminate the mismatch matching point and improve the matching accuracy,the SURF matching algorithm is combined with the limit constraint method to verify the feasibility of the improved algorithm.Based on the principle of binocular disparity,the known matching pairs are converted into three-dimensional coordinates,and the precise positioning of the door handle is achieved.For the problem of pose estimation of the door handle in the camera plane,this paper simulates the least square method and the minimum area rectangle algorithm,and analyzes and validates these two methods through experiments,and compares and selects the best calculation method.Through the establishment of a visual system experiment platform and training of neural networks,image collection and experiments were performed on the actual scene.According to the experimental results,it can be seen that in the industrial field,the designed neural network and 4-neighborhood segmentation method can meet the robot's need for door handle recognition.The principle of the SURF matching algorithm combined with the epipolar constraint algorithm was studied,and the matching accuracy of the door handle was improved.Correspondingly,the accuracy of the corresponding spatial three-dimensional coordinate results is also improved.The principle of least-square method plane fitting and minimum area rectangle method was studied.The results of the above two types of door handle pose estimation algorithms were compared and analyzed by experiments,and the optimal algorithm was selected.This research algorithm can solve practical problems very well,and it also has certain reference value for later research.