Research On Semantic Map Construction Technology For Dynamic Environment Of Workshop

With the rapid development of computer technology and artificial intelligence technology,the traditional manufacturing industry began to change to intelligent manufacturing based on information technology.Among them,the material handling system in the industrial field is an important part of the intelligent transformation,which provides a guarantee for the continuity of workshop production.Workshop material handling systems often use AGV as a transportation tool,but traditional AGV independent ability to perform a task is very limited,mostly along the guide path,not under the interference of dynamic objects,accurately positioning and build the map,and lack of semantic information awareness for the environment,and is difficult to perform high-level intelligence task.Therefore,in order to solve the AGV’s perception ability of workshop environment and improve its intelligence level,this paper carried out research on semantic map construction technology oriented to dynamic environment of workshop by using Simultaneous definition and Mapping（SLAM）technology.The specific research contents are as follows:（1）Semantic map construction system based on visual SLAM and deep learning is built.For AGV is difficult to build real-time map and the problem of semantic information for environment,with low cost and abundant information of depth camera as the sensor,use the visual SLAM complete tracking localization,the back-end optimization,selection of high real-time,high precision of instance network Yolact Edge get semantic information of the image segmentation,complete construction of the semantic map building systems.At the same time,aiming at the semantic recognition problem of workshop environment,the instance segmentation data sets of NUMERICAL control lathe,horizontal lathe and AGV are completed.（2）The dynamic target feature point elimination algorithm adapted to workshop environment is studied.Due to the interference of complex dynamic objects in workshop,AGV is easy to lose location information or get wrong pose estimation.In this paper,optical flow and motion consistency detection are used to achieve fast moving point detection,and then the exact contour segmentation of prior objects is completed based on case segmentation and the contour segmentation of unknown dynamic objects is completed by improved region growth algorithm.Moreover,constraints are added to the local moving prior objects to complete feature point filtering of dynamic objects.Finally,experiments based on open data sets and simulation workshops are completed to prove that the proposed algorithm can effectively eliminate dynamic targets in scenes and improve the robustness of semantic map construction system.（3）In order to realize semantic map visualization,navigation and extraction of key object information,semantic map construction based on object model library is studied.Firstly,semantic recognition,dynamic target screening and coordinate transformation are carried out based on key frame,and point cloud data is optimized.Secondly,the object model library is constructed,and the object is created and updated based on multiple screening mechanism and the model library is reduced based on ICP algorithm.Then the3 d semantic map based on Octomap and the 2d raster map suitable for navigation are constructed.Finally,semantic map construction based on public data sets and real scenes is completed,which proves the feasibility of the map construction method in this paper.