Research On Dynamic Vision SLAM Algorithm Based On Multi-focal Length Combination

Simultaneous localization and mapping(SLAM)technology is one of the indispensable key technologies in intelligent vehicles.It realizes map construction and environmental perception while realizing its own pose positioning.The image input of the standard binocular vision SLAM system is based on the camera with the same focal length lens combination,and its observation field of vision is consistent.Moreover,the traditional pure vision SLAM system assumes that the environment is global static,and the positioning and mapping in the real environment may be completely unreliable.In this thesis,a dynamic vision SLAM system which based on multi-focal length is proposed,and it overcomes the disadvantage that the standard binocular vision SLAM system can’t take into account the perception of long-distance and wide field of view environment,and also eliminates the influence of dynamic feature points on SLAM.The main research contents are as follows:1.Based on multi-focal length and Bouguet algorithm,a stereo calibration algorithm of multi-focal length binocular camera is improved.The stereo correction parameters of multi-focal length calibration can be used to stereo correct multi-focal length images.In this algorithm,instead of stereo correction of multi-focal length image,stereo correction parameters are used for the extracted image feature points.2.A feature extraction algorithm and feature matching algorithm are proposed based on image pyramid and quadtree algorithm.In the combination of 6mm and12.5mm lenses,the improvement ratio of this method is 85.36%;In the combination of12.5mm and 16 mm lenses,the improvement ratio of this method is 17.74%;In the combination of 16 mm and 25 mm lenses,the improvement ratio of this method is 31%.Experiments show that it can effectively increase the matching quantity and stability of multi-focal length stereo features.3.The self-made instance segmentation dataset is divided into five categories: road,automobile,human,motorcycle and cyclist.The instance segmentation network You Only Look At Coefficien Ts(YOLACT)segments a priori potential dynamic objects and road areas.The speed of YOLACT exceeds 25 frames on NVIDIA RTX 2060,and the mean average precision and the mean intersection over union values of self training detection weights are 48.3% and 44.4% respectively.4.Combined with multi view geometry,regional feature flow and relative position,three methods are used to identify whether the priori dynamic object segmented by the instance is a dynamic object.Finally,the feature points located on the surface and edge of the dynamic object are eliminated,and the completely static feature points are obtained to estimate their own pose,so as to realize the robust estimation of their own pose in the dynamic environment.5.The self-made SLAM dataset for multi-focal length and standard focal length binocular is used to verify the positioning accuracy of the multi-focal length dynamic vision SLAM algorithm in this thesis.The comparison experiment with the existing classical visual SLAM algorithm on KITTI dataset,and the comparison experiment between this algorithm and the existing classical visual SLAM algorithm on the selfmade dataset.The experimental results on KITTI dataset show that the localization error of the multi-focal length combined dynamic vision SLAM algorithm in this thesis is26.26% lower than that of ORB-SLAM3,and 7.30% lower than Dyna SLAM.In the self-made dataset,the positioning error of the multi-focal length combined dynamic vision SLAM algorithm in this thesis is 19.94% lower than ORB-SLAM3 and 6.17%lower than Dyna SLAM.The time performance of this algorithm is four times that of Dyna SLAM and the real-time requirement of 10 frames.