Research On Visual SLAM Based On Semantic Segmentation In Dynamic Environments

Most of the current visual SLAM algorithms are designed based on the assumption of static environment,and the robustness and accuracy are poor in the dynamic environment.The reason is that moving objects in the scene will cause mismatching of features in the pose estimation process and affecting the accuracy of positioning and mapping,so the dynamic environments needs to be processed.However,the SLAM algorithm dealing with dynamic environments has the problems of low accuracy and poor real-time performance.Therefore,the SLAM algorithm is designed for the dynamic environment,and the positioning accuracy and real-time performance of the SLAM system in the dynamic environments is improved.At the same time,the improved SLAM system obtains the semantic map of the dynamic environment.Firstly,the influence of dynamic environments on SLAM algorithm is analyzed from the perspective of pose estimation of mobile robot.Combining semantic segmentation and optical flow method,the moving object processing algorithm in dynamic environments is designed.Analyzing at the current moving object detection algorithm,the sparse LK optical flow is used to detect dynamic objects in the scene,and the image semantic segmentation of the image is used to obtain the accurate positional region of the potential moving object in the scene.Finally,the contour information of the moving object is extracted by combining the two information.In the subsequent positioning and mapping process,the feature points and spatial points of the moving object area are removed,and the influence on the positioning and construction of the SLAM system is reduced.The accuracy and real-time performance of the moving object processing algorithm are verified by experiments.Secondly,to finish the semantic segmentation,a convolution-based semantic segmentation network is designed.The SLAM algorithm is applied to the real environment,so the real-time requirements are higher.The semantic segmentation in moving object processing algorithm will affect the real-time of visusl SLAM algorithm.So a small and lightweight semantic segmentation network based on the convolution is designed by introducing dilated convolution,reducing the size of the network decoder,and using anti-pooling.The comparison between the relevant network shows that the network achieves the balance between accuracy and efficiency.Subsequently,the semantic map of the dynamic environments is constructed by combining the semantic information of the image and the SLAM algorithm.The moving object information obtained by the detection of moving object is used to remove the moving object part in the scene,and the construction quality of the SLAM system and the robustness in dynamic environments are improved.At the same time,the semantic fusion algorithm is introduced to optimize the continuously updated semantic information in the space map.Finally,the experimental platform is built,and experiments on pose estimation and semantic mapping in dynamic environments are carried out in the dataset and the actual scene.A large number of experimental results show that the SLAM algorithm designed in this paper can run well in high dynamic environments and the accuracy of pose estimation of mobile robots is improved.Besides,the static surrounding can be accurately mapped in the dynamic environments.