Research On Visual SLAM Algorithm Based On Semantic Segmentation In Dynamic Scenarios

Simultaneous Localization And Mapping(SLAM)means that without any prior information,the robot obtains the surrounding scene information through the sensor carried by oneself to obtain the surrounding scene information,and in the process of movement to estimate their own posture,and establish the environment perception model.The visual SLAM takes the camera as the sensor.By extracting and matching the features of the input images,the pose information is solved and the map is constructed.However,most visual SLAM algorithms are usually based on the assumption of static scenes.The drift errors caused by dynamic objects such as pedestrians cannot be eliminated,which limits the application of the algorithm to a certain extent.Moreover,these sparse maps constructed by traditional visual SLAM only contain some geometric information,which is difficult to meet the needs of mobile robots to complete advanced interactive tasks.With the rapid development of deep learning,visual SLAM combining semantic information can extract the information of objects in images more accurately,which has become a hot research direction in recent years.To solve the above problems,this paper studies semantic segmentation based visual SLAM algorithm in dynamic scenes,uses deep learning semantic segmentation network to extract semantic information,improves the accuracy of pose estimation while eliminating the interference of dynamic objects,and builds a three-dimensional map containing static information.Aiming at the problems of traditional visual SLAM in dynamic scene,such as the positioning accuracy is not ideal,a pose optimization algorithm in dynamic scene is proposed.At the same time of feature extraction,this paper uses PSPNet semantic segmentation network to extract the semantic information of the object,and uses the background feature points as the sample points to calculate the basic matrix,effectively combines the semantic information and geometric constraints to determine the true motion of the object,so as to eliminate the dynamic feature points and retain as many feature points as possible.When the camera lens is blocked too much by the dynamic object,the background feature points are too sparse,resulting in tracking loss at some point.In this paper,the camera motion is calculated from the motion of the blocked object,ensuring the continuity of pose estimation and improving the positioning accuracy of the algorithm.Sparse maps constructed for visual SLAM in dynamic scenarios are difficult to meet the high level interaction requirements of mobile robots,a3 D dense map construction algorithm in dynamic scene is proposed.In this paper,the key frames are further screened by coincidence detection to reduce the redundancy of point cloud data.Then combined with the object motion judgment,the dynamic object in the key frame is filtered out in the semantic mapping,and the missing background information is repaired statically.Finally,single-frame point clouds are spliced with pose information to build a static three-dimensional dense point cloud map.The experimental results of map construction show that the dense map constructed in this paper not only filters out dynamic objects,but also improves the efficiency of map construction.In order to verify the effectiveness of the algorithm proposed in this paper,the dynamic sequence of TUM open data set is used to test the object motion judgment and SLAM localization performance.Compared with the ORB-SLAM2 algorithm,the accuracy and robustness of the localization of the improved SLAM algorithm in the dynamic scene are verified.Finally,the static dense point cloud is effectively constructed in the dynamic scene.At the same time,the corresponding hardware experiment platform is built,and the real scene test is carried out to prove the effectiveness of the algorithm in this paper.