The Pose Estimation And Tracking Of Rigid Target Under Complex Environment

Pose estimation is a fundamental problem in computer vision and video measurement,and is often used in fields such as visual servoing and SLAM.Aiming at the pose estimation and tracking problem of the rigid body target of the carrier,this paper discusses the methods of geometric measurement and deep learning.In order to estimate the real-time pose of rigid body targets moving in complex environments,the pose estimation problem is modeled using perspective projection of linear features.Firstly,the pose estimation function is established by the matching measure function between lines and the linear perspective projection model.Then,in order to improve the adaptability of the algorithm,the pose parameters are robustly estimated,and finally solved by weighted least squares algorithm.Considering that the target motion has a certain continuity,for the image line extracted in real time,the fracture repair is performed according to the projection line of the previous frame.The joint pose estimation of multi-frame sequence images under the SFM framework can improve the pose estimation accuracy.The simulation and semi-physical experiments show that the benchmark method has higher precision and faster speed in the optimal solution of target pose in complex environments.Based on the pose estimation method of feature points and lines,it is usually studied that the rigid body is arranged in an arbitrary posture in three-dimensional space under the relationship of known 2D-3D points or line segment sets.In engineering practice,the rigid body target often moves on a certain spatial plane.Under this constraint,the existing point-or line-based pose estimation algorithm is improved by establishing a suitable coordinate system and transformation relationship,which will be calculated.The pose parameters were reduced from 6 to 3.Simulation and semi-physical experiments show that the series methods improve the robustness and speed of the solution to varying degrees.The traditional method of pose measurement is mainly based on the geometric matching of image feature matching and imaging.At the end of this paper,deep neural network is used for pose estimation.This paper divides the pose estimation problem into two steps: target detection and pose mapping.First,use Xception-like network as the backbone's Light Head RCNN for fast target detection,then use the lightweight Tiny-Net network for pose estimation.Apply this method to standard datasets and carrier targets.High precision and 28 fps speed are achieved in the indoor scene.