| With the rapid development of electronic technology and artificial intelligence,computer vision has made considerable progress in the past few decades.As a critical external sensor in this task,the camera has the advantages of low price,small size,and easy popularization and application.In addition,it can provide high-resolution color images at a relatively high frame rate under various weather conditions,and is widely used in robotic visual navigation,target tracking,intelligent transportation,and augmented reality.However,the coverage of a single camera is limited,and autonomous vehicles and robots generally install multiple cameras,and their wide field of view provides excellent potential for Omni-directional environmental perception.Therefore,the multi-camera system has received more and more attention in various fields.This article has carried out in-depth research on the multi-camera system’s calibration and relative pose estimation.The main work is as follows:(1)A calibration method for multi-camera systems is proposed based on calibration patterns without overlapping views between cameras.By relying on lidar to indirectly constrain the external parameters of the multi-camera system,this method overcomes the need for cameras to see the calibration object together.First,use the checkerboard to calibrate the internal parameters of each camera.Subsequently,a two-stage camera and lidar joint calibration algorithm based on the calibration board was designed to complete the external parameter calibration of the multi-camera system.In addition,the method has been verified on synthetic and actual data,indicating that the calibration method has high practicability and accuracy.(2)Two effective solutions based on the assumption of planar motion are proposed: the linear 6-point method and the minimal 3-point method.By modeling the multi-camera system as a generalized camera,the problem of relative pose estimation for indoor mobile robots is better solved with limited computational resources,while the motion of most indoor robots can be considered planar motion.Experimental comparisons with existing algorithms on simulated data show that the proposed algorithm has significant computational efficiency advantages and can be combined with a robust algorithm that removes outliers to achieve robust and efficient initial planar motion estimation.Finally,the test on real data further verifies the effectiveness of the solver.(3)A relative motion estimation algorithm for a multi-camera system based on affine correspondences is proposed,which uses a first-order approximation of the rotation matrix to compute the six-degree-of-freedom relative pose of the multi-camera system from a minimum of four affine correspondences.A multivariate system of equations is derived directly from the geometric constraints between the affine correspondence and the generalized camera model.The closed-form solution is then found by a linear method,and the accurate relative pose estimation parameters are generated.Finally,experiments on synthetic data and real image sequence datasets show that the affine solver effectively reduces the number of RANSAC iterations required for motion estimation compared to a point correspondence-based solver while maintaining comparable accuracy. |
