Research On Visual SLAM Algorithm In Indoor Dynamic Scene

Simultaneous Localization and Mapping(SLAM)algorithm is the basis for mobile robots to autonomously complete navigation tasks.SLAM obtains external environmental information through the sensors it carries,when the sensor is a camera,it is called visual SLAM.Although there are many excellent visual SLAM systems,most of the systems make static assumptions about the external environment,ignoring the impact of dynamic objects on the positioning accuracy and robustness of visual SLAM.Based on the ORB-SLAM2 system,we study the robustness of visual SLAM in indoor dynamic scenes,including three aspects:homogenized extraction of feature points,dynamic feature point filtering,and three-dimensional dense map construction.Improved the Qtree?ORB algorithm in ORB-SLAM2 to make the extracted feature points more uniform and reduce the extraction of low-quality feature points.In visual odometry,the ORB-SLAM2 algorithm is based on quadtree division to extract feature points.The extracted ORB feature points have a problem of over-uniformity,and many low-quality feature points are extracted.Based on this,first,the division depth of the quadtree is limited according to the number of feature points required for each layer of the image pyramid to reduce the number of iterations,and then the minimum value of Harris response value is limited to reduce the extraction of low-quality feature points.Finally,it is tested on the Mikolajczyk dataset and the TUM RGB-D dataset.The results show that the uniformity,the number of correct matches and the correct rate of matches can be effectively improved by the improved algorithm.The positioning accuracy of the SLAM system is tested,and the results show that the positioning accuracy of the SLAM system can be effectively improved by the improved Qtree?ORB algorithm.A dynamic feature point filtering algorithm based on semantic information and geometric constraints has been studied to reduce the impact of dynamic targets on the positioning accuracy and robustness of the visual SLAM system.First,YOLOv4 is used to detect a priori dynamic objects to filter out the priori dynamic feature points,the optical flow method is used to match the remaining feature points,combined with the epipolar geometric constraints to filter out the dynamic features point.Tested on the TUM RGB-D dataset,the results show that the dynamic feature points can be effectively filtered out.In the high dynamic sequence,the positioning accuracy of the SLAM system is significantly improved,which effectively improves the drift degree of the visual SLAM system,and the robustness is better;although the dynamic feature points can be effectively filtered out in the low-dynamic sequence,the improvement in positioning accuracy is not obvious,and it is not sensitive in the low-dynamic sequence.A three-dimensional dense map construction method is studied,and a static point cloud map and an octree map are constructed.In ORB-SLAM2,the map constructed with key frames has a high degree of overlap,excessive map redundant information and ghosts caused by dynamic targets,resulting in poor map readability.First,the key frames are checked for overlap,then the image depth information is detected,and finally the dynamic information in the scene is filtered out with the semantic information,and the static point cloud map and the octree map in the dynamic scene are constructed.The test results show that the constructed point cloud map and the octree map have little overlap,less redundant information,and the dynamic objects are basically filtered out,and the map is more readable.The overall performance of the improved visual SLAM system is tested in the actual environment to verify the performance of the improved SLAM system.The uniformity of feature points,the filtering of dynamic feature points,and the construction of three-dimensional dense maps were tested in three scenes:office,laboratory,and corridor.The test results show that the feature points extracted by the feature point extraction module in the actual environment have good uniformity and no over-uniformity occurs;the dynamic feature points in the actual environment are effectively filtered by the dynamic feature point filtering module;The point cloud map and the octree map constructed by the three-dimensional dense map building module can effectively filter out the dynamic information,and the map redundancy is small.