A Multi-sensor Fusion SLAM Framework Designed For Autonomous Driving

Real-time localization is one of basic technologies in the field of autonomous driving which can provide key information such as the position and the orientation of the vehicle in the world coordinate system.In order to quickly build a reliable real-time localization module,many companies and research institutions usually use high-precision integrated navigation equipment that combines satellite signals with inertial measurement units(IMU).This real-time localization solution has many disadvantages that cannot be ignored,the most serious disadvantages of which are poor environmental adaptability and high hardware cost.Therefore,more and more autonomous driving researchers have started to use the algorithm framework with SLAM(Simultaneous Localization And Mapping)as the core to support real-time localization needs.However,there are significant problems such as poor robustness and long calculation delays.At present,there is a lack of an open-source multi-sensor fusion SLAM framework for common hardware configurations of autonomous vehicles in this field.Based on above analysis,this paper designs a multi-sensor fusion SLAM framework for autonomous driving and proposes a new multi-sensor fusion mode:1)Proposing the concept of universal frame used to reduce errors caused by data synchronization and approximation;2)Using the redundant design of sensors to improve the robustness of entire system during real-time localization process 3)Introducing zero-speed bias calibration to optimize the estimation of bias of the IMU by making use of special driving conditions;4)Proposing the concept of time-sharing optimization to ensure that the calculation delay is maintained within an acceptable range.On the other hand,on the basis of the above framework,this article has proposed new solutions in three aspects of pre-integration constraints,visual constraints and lidar constraints:1)Proposing a new kind of pre-integral constraints that combines IMU and wheel speedometer;2)Making good use of road information in autonomous driving scenarios to improve the localization accuracy;3)Considering the fusion of feature point method and direct method to improve the robustness of visual SLAM technology in different kinds of environments;4)Incorporating visual semantic information and structural semantic information in the lidar constraints to improve the localization accuracy and calculation efficiency of the lidar SLAM technology.The SLAM framework proposed in this paper has been applied to autonomous vehicles as a real-time localization solution and the mileage has exceeded one thousand.This framework basically meets the efficiency requirements and accuracy requirements of autonomous driving.