Design And Implementation Of Multi-Sensor Simultaneous Localization And Mapping System

Currently,autonomous driving,as an intersection of computer science and automotive industry,has a broad market prospect and is rapidly becoming one of the hot areas in artificial intelligence.Autonomous driving contains key modules such as localization,perception,decision making and navigation.Simultaneous Localization and Mapping（SLAM）can effectively combine the advantages of multiple sensors to estimate the vehicle position,construct the surrounding environment,and provide a key reference for building a high-precision map containing various road condition information.In summary,accurate and reliable state estimation and map construction are important components of autonomous driving systems.The current mainstream methods focus only on geometric feature alignment methods in static scene tasks.On the one hand,geometric feature-based alignment is very susceptible to interference from dynamic objects,which leads to accuracy degradation.On the other hand,with the development of deep semantic segmentation networks,we can easily obtain semantic information from point clouds in addition to geometric information.Semantic features can be used as a supplement to geometric features,which can improve the accuracy of odometry and loopback detection.Semantic information can also filter out dynamic objects in the point cloud data,such as pedestrians and vehicles,which can lead to residual images in the generated maps or map-based localization failures.Through the above analysis,this paper improves the front-end odometry and loopback detection modules based on the LIO-SAM framework,and the main contributions of this paper are as follows.（1）In this paper,we optimize the semantic information provided by the SPVNAS network and combine it with the LIDAR front-end odometer.The algorithm can still perform accurate alignment of the system in the experimental scenario with a large number of moving objects.（2）By combining semantic information,this paper proposes a semantic-assisted Scan Context method and uses semantic ICP to verify its results and apply it as a loopback detection strategy for the system.The algorithm makes the system focus more on static scene matching during loopback detection,which can correct the cumulative error of the poses and improve the global map building effect.（3）In this paper,we collected Jilin University campus data in various scenes using laboratory sensors,and conducted multiple correlation control experiments on the classical autopilot dataset KITTI and Jilin University campus data.The experimental results show that our method outperforms pure geometric methods,especially in dynamic scenes,and has good generalization ability.