Research On Semantic Information-based Visual SLAM Approach In Substation Scenes

As the construction of smart grid continues to deepen,manual inspection is challenged by heavy workloads,insufficient manpower and high risks.In response to these problems,the power grid corporation vigorously promotes the development of unmanned and intelligent substations.In this process,inspection robots have been widely used,and the localization and mapping of robots are the key technologies for achieving their autonomous inspection.Simultaneous localization and mapping(SLAM)is an important way to solve the localization and mapping.Currently,SLAM applied to substations mainly uses laser SLAM,whose high cost limits the promotion and application of robots.Moreover,the constructed maps lack semantic information,which is not conducive to robots to perform environment perception and execute more advanced tasks.In order to overcome the above problems,this thesis carries out the research on the visual SLAM method based on semantic information.The thesis mainly includes the following research contents.(1)A real-time semantic segmentation method based on boundary information guidance.To address the complexity of the substation scene and the boundary blurring and small target omission in semantic segmentation,this thesis introduces a boundary branch to obtain boundary information of the image to enhance the network’s attention to boundaries.In order to speed up the inference,this thesis improves the real-time performance of the network by sharing features,reducing the number and size of feature map.In addition,this thesis uses the improved pyramid pooling module and attention fusion module to extract contextual information and fuse features.Experiments show that our method achieves high segmentation accuracy and inference speed on both the Cityscapes dataset and the substation scene dataset,and effectively improves the segmentation of the network in boundary ares and small targets.(2)A map construction method integrating semantic information.On the one hand,to solve the pose errors caused by dynamic feature points in visual SLAM,this thesis proposes a dynamic point removal method that combines semantic and motion constraints.The algorithm converts the segmentation results into binary masks,and then combines motion constraints based on optical flow and dynamic consistency to remove dynamic points.On the other hand,to address the problem of lacking semantic information in traditional maps,this thesis proposes a method that integrates semantics into octree map to generate semantic maps.This method projects 2D semantic information onto a 3D point cloud and then produces a semantic map by sparsifying the point cloud using octree grids.(3)Semantic mapping and localization in substation scenes.This thesis sets up a hardware and software platform,and tests the above algorithms in a substation scene.The results show that the SLAM method proposed in this thesis can achieve accurate mapping and localization in complex substation scenes,and provide rich semantic information for the environmental perception of inspection robots.