Research Of Loop Closure Detection For Sparse Semantic Mapping

Compared with traditional maps consisted by visual feature and dense pointcloud,a semantic map contains the category information of mey objects in the surrounding environment,which is more recognizable than visual features and closer to human's comprehention towards the surrounding scenes,providing more direct evidence for autonomous path planning,navigation and human-computer interaction.On the other hand,semantic information is a kind of powerful tool for camera pose estimation,back-end optimization and scene description for visual SLAM(Simultaneous Localization and Mapping)problem.Different from dense semantic map which focus on details of the scenes,sparse semantic maps contain semantic information as map points with categories in a sparse semantic map.For sparse semantic mapping,in order to solve the problem of loop closure detection,this thesis makes use of RGBD sensor,YOLOv3 object detection algorithm is adopted to obtain semantic information in scenes,and the effectiveness of proposed algorithm is verified by experiments,sparse semantic maps also have been established.Aiming at the problem of false positive detection caused by current loop closure detection methods in visual SLAM,this thesis proposes a method of scene description and similarity computing based on object detection and unsupervised learning,and then performs loop closure judgement.Density clustering algorithm DBSCAN is used to correct wrong and missing object detection,so as to generate semantic nodes corresponding to semantic information,and compute the confidence of the semantic nodes according to clustering results.Making use of semantic nodes in keyframes,local semantic topologies for keyframes are established for scene description by analyzing relative positional relationship between the semantic nodes in the camera coordinate.When computing similarity of different scenes,semantic nodes are matched according to SURF features and object category,and the transformation of corresponding edges in different semantic topologies is calculated to obtain the similarity between the scenes.Judgement of loop closures is performed according to the changes of similarities between consequent keyframes.Experiments on benchmark datasets prove that object clustering effectively improves the accuracy of loop closure detection in indoor scenes.Compared with algorithms which are barely based on traditional visual features,the proposed algorithm can achieve loop closure detection with a higher accuracy.Based on above loop closure detection algorithm,this thesis combines a visual SLAM system framework ORB-SLAM2 to carry out global optimization experiments,and combines the YOLOv3 object detection results to construct a sparse semantic map.The experimental results show that the global optimization effect based on the loop closure detection results that came up with the proposed algorithm get a better result than the traditional method,which effectively reduces the error of keyframe trajectory.In addition,unsupervised learning method are used to improve the semantic map points in sparse semantic maps,the maps are effectly optimized by a method which combined density clustering and prototype-based clustering.