Research On Vision-based Semantic SLAM Algorithm In Dynamic Scenarios

Simultaneous Localization and Mapping(Simultaneous Localization and Mapping,SLAM)obtains environmental information through the carried sensors to jointly estimate the movement of the robot and the map of the surrounding environment.Most SLAM systems assume that the robot runs in a static environment.However,the actual running scene of the robot contains a large number of moving objects,which will seriously affect the accuracy of the system positioning and even cause the failure of the system positioning.To solve these problems,a stereo vision semantic SLAM method in outdoor dynamic scene is proposed in this paper.The specific content is as follows:Firstly,in order to solve the problem that the feature points extracted from dynamic objects affect the positioning accuracy of the system,a method of using image semantic segmentation information to eliminate dynamic points is proposed.The convolutional neural network is used to segment the image and obtain the category information of pixel points.Based on this,the dynamic attributes of feature points are defined and removed.The static feature points are used for matching,and semantic attribute matching is added in the matching process of feature points,so as to restore the relatively accurate camera pose.Experimental results prove that the algorithm effectively reduces the impact of dynamic objects and improves the positioning accuracy of the system.Secondly,aiming at the problem that the current motion state of the object cannot be judged only by using the semantic information of the image,which leads to the loss of the dynamic object without moving feature points in its current state,this paper proposes a method combining semantic segmentation and multi-view geometry to eliminate the moving object and recover the feature points on the object mistakenly removed by the semantic information.In order to determine the current motion state of the feature points,the camera pose was estimated based on the constant speed motion model and the spatial position of the feature points was recovered.The spatial velocity of the feature points was calculated and the feature points were selected based on the semantic information.In order to solve the noise interference when calculating the characteristic point velocity,the mean filtering window is designed to filter out the noise points.The experimental results show that using the combination of scene flow and semantic information,the system can not only eliminate the influence of moving objects on positioning,but also restore the feature points that were mistakenly eliminated by semantic information,and improve the positioning accuracy of the system.Finally,to solve the problem of erroneous or incomplete culling of dynamic points,a method of selecting dynamic spatial points based on long-term observation map points was proposed,and semantic map was constructed.In terms of eliminating the influence of dynamic objects,the occurrence times of space points in continuous frames were recorded,and the optimization times were taken as the basis for the stability of the feature points.With the increase of the observation times,the confidence of potential dynamic points was gradually increased and participated in the optimization,so as to improve the positioning accuracy of the system.In terms of semantic map construction,in addition to obtaining semantic segmentation images,the probabilities of different categories of each pixel point are recorded to construct semantic map information.Bayes is used to update the map points as the semantic attributes of observation change,and finally the semantic information of spatial map points is determined by the maximum probability category.Experimental results prove that the proposed algorithm can distinguish dynamic points from static points,improve the accuracy of system positioning,and finally realize the construction of semantic maps in large-scale environments.