Research On Localization And Mapping Algorithm Of Unmanned Vehicles Based On Multimodal Sensor Fusion

Simultaneous Localization and Mapping(SLAM)is one of the core technologies for unmanned vehicles to realize autonomous navigation,path planning and other functions in unknown environments.Due to the complexity and variability of the actual driving environment and the continuous improvement of intelligent requirements,the traditional SLAM algorithm based on a single sensor is difficult to meet the needs of the unmanned vehicles for robust positioning and environmental map construction,let alone support their high-level understanding of the driving environment.Therefore,this paper proposes a localization and mapping algorithm LADVI-SLAM based on multimodal sensor fusion.On the one hand,the performance complementarity is achieved by fusing different sensors to improve the robustness of the algorithm in the face of unknown driving environments.On the other hand,neural network is used to make up for the perception limitations of traditional SLAM,and add semantic information to the map while weakening the interference of dynamic objects.The main research contents of this paper are as follows:(1)In order to effectively combine the accurate depth measurement of Li DAR with the powerful texture perception advantages of vision,a depth-vision perception algorithm based on the fusion of Li DAR and vision is proposed.Firstly,the heterogeneous data conversion from point cloud to depth image is completed according to the imaging model of the camera,and then visual constraint is used to achieve dense alignment between pixels.At the same time,in order to solve the depth loss problem caused by the sparsity of the point cloud,the adaptive expansion kernel is used to complete the rapid densification of the sparse depth image,and the final depth-vision perception result is obtained after smoothing the image through bilateral filtering.The algorithm is coupled to LADVI-SLAM’s data fusion preprocessing module to improve its adaptability and robustness in the face of different driving environments.(2)Due to the presence of dynamic objects,traditional SLAM algorithm suffers from problems such as decreased positioning accuracy and repeated modeling during mapping.To address these problems,a dynamic object detection algorithm based on the joint constraints of geometric consistency and semantic information is proposed.Firstly,based on the Bi Se Net V2 neural network,a priori motion probability model for different targets in the driving environment is established,and then the true motion state of the internal feature points of potential moving targets is judged using epipolar geometry.Through the joint constraints of the two,the object-level dynamic object detection is completed.On the basis of comprehensively removing dynamic feature points,a depth-vision-inertial odometry for dynamic driving environment is further realized as the front end of LADVI-SLAM,thereby providing more reliable positioning results for unmanned vehicles.(3)Aiming at the problem of traditional point cloud maps lacking color,texture,and higher-level semantic information,a dense point cloud map construction algorithm based on3 D Normal Distribution Transformation(3D-NDT)matching is proposed to support the map construction function of LADVI-SLAM.Firstly,a high-density 3D point cloud carrying color information is generated through the depth-vision perception algorithm,and the dynamic object detection algorithm is used to give the point cloud semantic information while removing the point clouds of moving objects.Then,a globally optimal point cloud map is constructed using a scan-to-map approach.Finally,the map is optimized based on spatial structure and weighted voting mechanism.(4)Based on the public dataset of KITTI and real vehicle scenes,the LADVI-SLAM algorithm proposed in this paper is verified by experiments.The results show that compared with the comparison algorithm,LADVI-SLAM can better adapt to different types of driving environments,have higher positioning accuracy,and the constructed color point cloud map and semantic map have better global consistency on the basis of containing more information.