Research On SLAM Based On Multi-sensor Fusion In Dynamic Scenarios

Simultaneous Localization and Mapping is considered to be a basic ability of intelligent robots.In recent years,visual SLAM is becoming more mature.Plenty of v SLAM have not only great trajectory accuracy but also high precision mapping effects in specific scenarios.However,in certain scenarios,these systems are not robust,such as highly dynamic scenes dominated by dynamic objects in the environment.These systems cannot perceive dynamic objects,which greatly affects the robot's position accuracy and navigation effects.In order to perceive dynamic objects,some systems make the robot understand the environment in a true sense,making the system sacrifice real-time.Today how to make the robot system detect moving objects in a dynamic environment in real time,so as to complete higher-level tasks,has become a hotspot of research.In allusion to the existing system lack of robustness in the dynamic scenes,we improve on the classic v SLAM framework,ORB-SLAM2.The improvement ideas are as follows.Aiming at low trajectory accuracy of classical v SLAM system in dynamic scenes,we fuse the encoder sensor in the system and use the encoder as the auxiliary sensor.In the robot tracking module,we use the encoder model to replace the constant velocity model in the classic v SLAM system.By combining the encoder with the external parameter matrix calibrated by the robot,the system provides a more accuracy initial value for solving the pose between frames.At the same time a dynamic keypoints elimination algorithm base on encoder and probability statistics is proposed,which can effectively suppress interference of dynamic objects on robot trajectory tracking and improve the trajectory accuracy of the system.The Experiments show that the trajectory performance of the system is significantly improved compared with the mainstream dynamic SLAM,DS-SLAM,in high dynamic scenarios.In terms of ATE,the RMSE improvement values can reach up to 36.39%,while the optimization of trajectory does not sacrifice the real-time performance of the system.Our system take less time to process each frame than DS-SLAM and its performance is improve by 47.24%,which meets the demand of real-time.In the response to the poor quality of mapping in dynamic scenes,we incorporated semantic segmentation into the system.For the low accuracy of segmentation in high dynamic environment,we propose a semantic completion algorithm,which is able to effectively complete semantic tags on dynamic objects.We use geometric consistency algorithm to determine whether the semantic area is a dynamic object,and eliminate dynamic pixels in key frames.The system converts the key frames carrying static semantic information into point cloud by using depth values to project them to world coordinates and finally transforms the point cloud into a static octree map with semantic information.In addition,this article also explores other methods that v SLAM system combines with deep learning.We propose a algorithm for removing dynamic feature points between frames base on YOLO v3.The algorithm selects the detection frame from ordinary frame at a specific frequency,performs YOLO v3 target detection on detection frame to find out the dynamic objects,and uses multi-view geometry to reproject the detection result onto the ordinary frame,eliminating the dynamic feature points in the ordinary frames,which effectively reduces the time for target detection on normal frames.