Research On Pose Measurement Based On Vision Of Autonomous Target

With the rapid development of image processing technology,visual pose measurement has attracted wide attention because it can estimate the pose of the target in real time without environmental prior information.The complexity of unfamiliar environment has brought great challenges to the pose measurement system.Research on autonomous pose measurement with high accuracy,good real-time performance,and strong robustness has become an important direction in the field of computer vision.Based on the existing feature-based binocular visual pose measurement method,this paper mainly focuses on the problems of the difficulty of real-time target tracking,high complexity of optimization calculations and inconsistent scales,thus some new methods are proposed.The main work is as follows.Firstly,a feature extraction algorithm based on contrast limited adaptive histogram equalization is proposed.The robustness of the algorithm is improved by histogram equalization with adaptively limited contrast,where the accuracy of real-time pose tracking algorithm between images based on data association is guaranteed.In addition,this paper proposes an adaptive selection of motion model algorithm for pose tracking by judging the current motion state of the object,which enhances the stability of the system,and designs an appropriate key frame generation strategy for subsequent local optimization.Secondly,a local pose optimization algorithm based on sliding window is proposed.With the localization represented as a sliding window,the pose of local key frames and 3D map points is optimized non-linearly.The sliding window is used to store the local key frames and update the data association relationship,in which case the key frames and 3D points with poor correlation and long-time interval are truncated,and the error transmission through pose measurement between adjacent images is eliminated as well,thus the performance of pose optimization is highly improved.Besides,a strict key frame selection strategy is used to eliminate redundant data and save computing resources.Additionally,data association based on sliding windows is independent of nonlinear optimization,avoiding feedback loops and thus preventing the feedback propagation of error.Thirdly,a loop detection algorithm with image similarity detection based on inverted index is proposed.In the process of long-term target motion,inverted indexes of visual feature vector words are used to identify repetitive scenes,which accelerate the speed of identifying closed-loop candidate frames from the key frame database during loop detection.According to the consistency of the two ends of the closed-loop,the global key frame pose is corrected to eliminate the accumulated error and improve the accuracy of the pose measurement results.Furthermore,loop detection provides a relocation choice in the case of failure of pose tracking,which enables the measurement to be carried out continuously when the matching between adjacent frames fails and enhances the stability of measurement.Finally,a pose measurement framework combining loop detection,pose tracking and local pose optimization is built.With experiment performed on public data sets,the performance of the proposed method is analyzed.The experimental results show that the measurement of the proposed methods is proved to be more accurate and robust to the change of ambient light compared with the existing mainstream methods.Our methods save the computing resources of pose optimization and achieve the balance between measurement accuracy and efficiency.