Simultaneous Localization And Mapping Of Robot In Short-term And Long-term Dynamic Environments

In recent years,with the rapid development of artificial intelligence,the field of robots has gradually integrated into our daily life,making our life more convenient.The development of mobile robots could not be separated from a key technology: simultaneous localization and mapping(SLAM).Most of the existing SLAM technologies are realized by combining vision,and the current visual SLAM technology is based on an important assumption: the environment of the robot is static.However,the actual environment in which the robot operates is dynamic,which can be divided into two situations.One is that there are highly dynamic objects in the robot operating environment,such as pedestrians,vehicles,etc.,which can be called short-term dynamic environment due to the short time of environment change.The existence of highly dynamic objects will affect the localization algorithm.The other is the change of illumination and angle of view in the robot operating environment,which can be called long-term dynamic environment due to the long time of environment change.The change of illumination and angle of view in the environment will affect the loop closure algorithm.By combining the advantages of deep learning in image processing,this paper designs a visual SLAM algorithm which can adapt to the dynamic environment.The current mainstream visual SLAM framework is used,which is divided into three parts: tracking,mapping and loop closure detection.In the tracking thread,extract orb features from image for matching,achieving the pose estimation and optimization in robot motion.In the mapping thread,we project the feature points of interest to the robot into the map,and build a more consistent map with the environment.In the loop closure detection thread,we detect whether the robot returns to the previous place by judging the image similarity to further eliminate the accumulated error.According to the characteristics of the short-term dynamic environment,this paper integrates SSD target detection algorithm based on Mobile Net V2 framework in the tracking thread of the SLAM algorithm.Through SSD target detection algorithm,the SLAM system identifies the area where the dynamic object is located,and then proposes particle filter algorithm to cluster and track the dynamic object,and improves the positioning effect of robot by eliminating the dynamic features that affect the positioning.The experimental results show that the localization accuracy of the SLAM algorithm constructed in this paper is consistent with or even better than the current best SLAM algorithm in the static scene,and the localization error is greatly reduced in the dynamic scene.According to the characteristics of the long-term dynamic environment,this paper integrates the lightweight convolutional neural network HOG-NET in the loop closure detection thread of SLAM algorithm,extract the deepth gradient direct descriptor from the image and store it in the database.We identify and match the position by calculating the Euclidean distance between the descriptors.At last,we design the loop closure detection optimization algorithm based on HOG-NET.The experimental results show that our method is better than the traditional bags of binary words model in the accuracy and query time,and has robust loop closure recognition ability in the dataset and the actual scene.