Research On Visual SLAM Based On Dynamic Environment

SLAM(Simulated Localization and Mapping)is widely used in the fields of autonomous driving,virtual reality and robot path planning.In the unknown and complex environment,SLAM completes the established tasks by constructing sparse feature maps.However,the traditional visual SLAM is based on the assumption of static environment,and the accuracy and real-time matching of feature points are poor,which makes the robot unable to achieve accurate positioning and low efficiency,and it is difficult to play a good role in the real dynamic environment.In addition,relying on traditional sparse images to extract features and matching feature points to estimate inter-frame data to achieve closed-loop detection can not make the robot complete advanced tasks containing semantic information.Aiming at the above problems,this paper studies visual SLAM based on semantic information in dynamic environment.The main contributions of this paper include:Firstly,Aiming at the problem that traditional SLAM uses brute force matching method to match feature points,which leads to low efficiency and low accuracy,a method of accelerating SLAM feature point matching by using target detection to establish semantic information is proposed.Aiming at the problem that the target detection algorithm is difficult to process continuous multi-frame images in a real-time visual SLAM system,the SSD(Single Shot Multi Box Detector)target detection algorithm is improved.Many experiments show that the improved SSD algorithm improves the correct recognition rate by 25% and reduces the matching time of feature points by 20%.Secondly,Aiming at the problem that the visual SLAM is disturbed by some abnormal points in dynamic environment,which leads to the low positioning accuracy and the low quality of the map,the method of deleting abnormal points is improved,the collected image is semantically divided and expanded,and the map is built and stored by using octree structure.Finally,this paper builds a SLAM intelligent robot system in a real scene,and verifies the availability and effectiveness of the above methods,and then experiments and verifies the overall performance of the robot system.The system realizes autonomous positioning,semantic perception and pose estimation,and at the same time constructs an accurate three-dimensional semantic map.Experiments show that the SLAM system based on the improved algorithm in this paper has better performance in accuracy than other SLAM systems,and the accuracy has increased by more than 5%,which verifies the practicability and accuracy of the visual SLAM in dynamic environment.