Research On Semantic SLAM Of Robots In Dynamic Scenes Based On Deep Learning

The real-time localization and mapping(SLAM)refers to the process that a robot estimates its posture and builds its environment map only through its own sensors in a strange environment.It is a prerequisite for many robot application scenarios,such as path planning,collision free navigation,environment perception,etc.Vision slam refers to the location and mapping method with camera as the main data acquisition sensor.Compared with other sensors such as laser,camera has many advantages such as low cost,small size,and so on.Moreover,vision provides rich texture information for robot processing scene,which is an important basis for generating environment semantics.Although there are many successful open source solutions,visual slam still has many shortcomings.First of all,the current SLAM Based on either feature point method or direct method only constructs the location of spatial point cloud,which does not show the visual meaning represented by the point cloud.Secondly,it is also because of the direct matching processing in the process of data association,which does not consider the semantic information of the environment,which leads to the greatly reduced calculation accuracy of the system trajectory when there are high dynamic objects in the environment,and even leads to the failure of system mapping.In order to solve the above problems,this paper proposes a method to construct a threedimensional semantic map by using deep learning to recognize the semantic information of the environment and combining the optical flow method to identify the dynamic objects.The semantic map constructed by this method can be either a point cloud map with dense interaction or an octree map used in the navigation process of indoor robots.In order to realize this process,the whole system adopts the multithread processing method,the input of the system is color map and depth map,and the semantic segmentation thread and slam process run at the same time.Firstly,the case segmentation network is used to segment the key frames of color image to establish the prior semantic information.At the same time,the slam process optimizes the method of solving the pose according to the dynamic information,and carries on the threedimensional semantic optimization,establishes the dynamic object information database,realizes the dynamic update of the semantic map.In order to facilitate the robot navigation process,the dense point cloud map can be transformed into the Octree map,and finally built Set up semantic map without dynamic object interference.The results show that the proposed method can effectively eliminate the impact of dynamic objects on the slam process,greatly improve the accuracy of robot traj ectory calculation in the process of real-time positioning and map building,and build a reliable three-dimensional dense semantic map or octree semantic map.