A Study And Implementation Of SLAM System Algorithm Based On Semantic Messages

SLAM is the key for mobile robots.SLAM uses sensors to detect environment around mobile robots,builds the 3D map and localize the positions of mobile robots.Recently,SLAM is developed so fast that it can work well in a large environment without dynamic objects.However,SLAM still face three problems.The first problem is that SLAM is easy effected by dynamic objects,and it can't work well in a environment with dynamic objects.The second problem is that SLAM can't obtain semantic messages from environment,so it can only generate the 3D map only containing geometric messages which can't fulfill the mobile robots' requirements for executing more advanced missions.The third problem is that Octomap is widely used in the field of building maps,but it still can't build the 3D map in real-time with a high resolution because of the low efficiency of building maps.In order to solve these three problems,this paper proposes the semantic SLAM based on state-of-the-art in the field of Featured-Based SLAM,ORB-SLAM.We do improve ORB-SLAM in five ways.First,we modify the tracking thread to let SLAM make better use of semantic messages.Second,the detection module based on deep learning methods is integrated into the Local Mapping threads of ORB-SLAM,and CRF(Conditonal Random Field)is used to optimize the result of detection module.Third,we combine semantic messages gathered by SLAM to generate objects in the map when SLAM tracks.Objects generated by SLAM contain semantic messages,point clouds and the position in the world coordinate system.Fourth,we build the relationship between objects and keyframes to integrate concept of semantic into the framework of ORB-SLAM.Fifth,we convert the type of the maps from point clouds to Octomap.And we modify Octomap to improve the speed of building maps.We evaluate our semantic SLAM in three different ways(Localization,Object Detections and Mapping)respectively.Specifically,the accuracy of localization and the map speed are evaluated on TUM dataset.From a comparison of the root-mean-square error(RMSE)with ORB-SLAM and original RGB-D SLAM,we found that the RMSE of semantic SLAM is close to the original ORB-SLAM in the static scenes,while semantic SLAM is much better than the Original SLAM and RGB-D SLAM in dynamic scenes.In the field of object detection,we demonstrate the efficacy of object detection through quantitative evaluation in an automated inventory management task on a real-world dataset recorded over an office.In the field of mapping,we compare the speed of the improved Octomap and the speed of original Octomap.The result shows that the improved Octomap is twice faster than the original Octomap.