Research On Key Techniques Of Adaptive Compensation Visual Odometrv In Dynamic Scenarios

In dynamic scenarios,dynamic participants in the scene break the static assumptions of the traditional Visual Odometry(VO)algorithm.It is common practice to remove dynamic participants and only use static participants as motion references.When the proportion of dynamic participants in the scene is too high,and then the participants with dynamic semantic labels in the scene are rudely removed,the number of reference feature points for pose estimation is drastically reduced,the robustness of the system and the accuracy of pose estimation.It will be much lower.In order to better solve the influence of scene dynamic degree on system robustness,this paper proposes an adaptive compensation VO system.The main research contents and innovations of this paper can be summarized as follows:1)Through the analysis of the principle and research status of Visual SLAM(Visual Simultaneous Localization And Mapping,VSLAM),and theoretically deducing the three-dimensional coordinate system transformation,camera calibration and depth information registration principle,this paper constructs the self.A sparse static feature map construction model adapted to compensation,which makes the compensation process independent thread and has greater flexibility,so that the system can better adapt to dynamic scenes.2)Through the in-depth analysis of the feature point preprocessing process in dynamic scenes and the experimental verification of the semantic segmentation effect of Mask R-CNN,this paper proposes a layered extraction fusion framework to realize the equi-probability feature points from global to local.Homogenization sampling,combining scene mapping multi-level weights and self-organizing spatio-temporal priority information,through the experimental analysis of layered extraction fusion frameworks at different levels,shows that layered extraction fusion framework improves ORB feature point matching accuracy and enhancement system It plays an important role in the generalization ability under different dynamic degree scenarios.3)Through the in-depth study of the moving target detection algorithm and the theoretical analysis of the motion compensation principle,combined with the needs of the system structure,this paper proposes a candidate pixel detection and extraction algorithm based on motion compensation,based on the semantic segmentation of Mask R-CNN network.The motion compensation and constant velocity motion model are used to recover the dynamic participant pixels in the previous frame,and the candidate pixels are detected according to the coincidence threshold,which greatly improves the accuracy and robustness of the motion detection.The experimental analysis of different coincidence thresholds shows that the coincidence threshold setting of the system has a weak effective interval,which is not sensitive to the change of the artificial setting value,indicating the rationality of the combination of the algorithm and the system.4)Through the experimental analysis of the motion compensation effect of the target pixel and the in-depth study of the image mask correction.In order to further optimize the discrete pixels after motion compensation,the accuracy of candidate pixel detection extraction is improved.For the modeling and processing of discrete pixels,a target image mask correction algorithm based on skeleton divergence is proposed.Firstly,the edge of the whole discrete pixel region is extracted and smoothed by the bilateral filtering,and then the contour information of the discrete discrete pixel region is extracted and optimized by using the edge information with higher confidence and Delaunay triangulation.The skeleton is optimized as the center of the local optimization,and the region is merged by the edge guidance.In the pixel region merging process,the ideal merging strategy is sorted by the TOPSIS method and the point cloud regularization is used for verification.In the regularization verification stage,the iterative simplification is carried out by means of increment and hierarchical clustering,so that the point cloud sampling is concentrated in the high curvature region and the surface deformation model is established,and the optimal suturing strategy is screened by the size of the edge stitching error.Experiments on single-objective and multi-target pixel regions,and experimental comparisons between different algorithms of the same sequence and different sequences of the same sequence also show that the proposed algorithm can overcome the excessive shrinkage and expansion of discrete pixel regions,and the experimental results are better than other existing other algorithms.The overall experiment of the system is carried out on the TUM RGBD dataset[79].Compared with many excellent visual mileage calculation methods,the absolute trajectory error and relative trajectory error of camera motion are obtained in most scenes with different dynamic degrees.Significantly reduced,the proposed algorithm demonstrates more robust robustness and higher accuracy in dynamic scenarios.