| Under the development trend of the interconnection of everything,unmanned carrier systems play an increasingly important role in aerospace,weapon manufacturing,data acquisition,industrial processing and other fields.The simultaneous positioning and mapping technology is an important basis for the role of unmanned carriers.As the most widespread sensor,vision sensors own unparalleled advantages in information collection.Many algorithms focus on studying how to abstract robust features from massive amounts of visual information.The change of the feature in the carrier coordinate system reflects the change of the carrier in the world coordinate system.There are two methods that perfectly recover the optimal estimate of the carrier motion from the changes of abstract features.Based on a strict comparison of the twodimensional image feature extraction algorithms and the simultaneous positioning and mapping back-end optimization algorithms,this thesis proposes a visual measurement and positioning algorithm based on a monocular sensor,which achieves the construction of a globally consistent map and a posture recovery with relative error less than 10%.The main research contents of this article are as follows:(1)Through the study of imaging models,a mapping from a point in real space to a point in pixel space is established,and each intermediate quantity in the mapping process is accurately described using different coordinate references.Compensation of radial distortion errors and tangential distortion errors is achieved through a unified correction model.In the experimental part,accurate sensor parameters is obtained.(2)In the process of researching the image feature extraction algorithms,a single image with unique dimension is expanded into a tower-like structure with continuous dimensions,and robust image features and descriptors are obtained.The epipolar geometry and the PnP method are studied,and the conversion between feature motion and sensor motion is realized.The performance of the three extraction algorithms is compared from the perspective of the number of feature points and the process time.In two sets of experiments,it is concluded that the efficiency of the ORB algorithm is at least 3.38 times that of the suboptimal algorithm.(3)Probability theory is introduced into the problem of sensor motion estimation,and the state reliability estimation is realized.The different optimization methods in back-end optimization are studied,and the conclusion that the biggest difference between the two methods lies in Markov properties is obtained.An evaluation function for judging the performance of the optimization process based on different data before and after optimization is proposed.In the experiment,the superiority of the nonlinear optimization method is proved by the index that comes from the result of data after optimization minus data before the procedure under this function.(4)Driven by data from different sources,experiments are performed on the proposed vision measurement and positioning system.In the stability experiment,the single trajectory results converge,and the corresponding trajectory results in 14 experiments tend to be consistent.In the validity test using the data set,the umeyama algorithm is used to align the estimated trajectory with the real trajectory for the problem of scale degradation.In the three groups of experiments,the maximum relative error is 0.058 m,and the maximum absolute error is 0.051 m.In the validity experiments using the recorded data,the relative trajectory error of the circular trajectory is 7.56%,and the relative trajectory error of the rectangular trajectory is 9.32%. |
