Study On Optimization-based Multi-sensor Fusion Motion Estimation Approach Of Intelligent Vehicle

Precise and drift-free motion estimation can help intelligent vehicles better understand the relationship between the body and the environment,and determine the advanced level of intelligent vehicles in localization,path planning,and obstacle avoidance.It is an essential and fundamental technology for autonomous driving.A larger number of current algorithms mainly focus on single-sensor methods,but in fact single-sensor methods such as laser-based or vision-based methods have been proven to be inadequate to cope with the challenges faced by autonomous driving in complex environment scenarios.Meanwhile,fusion algorithms that combine vision with inertia or laser have been proposed.However,due to the characteristics of these sensors,they measure the relative pose relative to the starting point in their own coordinate system instead of the absolute pose in the global coordinated system,therefore,these algorithms still accumulate drifts over time.In response to this phenomenon,this paper proposes an optimization-based multisensor fusion approach that can support the fusion of multiple sensors(including stereo camera,Li DAR,IMU,GPS,etc.).The core idea of the fusion method is to construct a globally unified pose graph through a dual-layer optimization strategy,and integrate complementary information from different sensors in the graph,so that high precision and global drift-free effects can be achieved.First,the system starts with data processing of each sensor,including recovering the visual structure of the camera from images,aligning the inertial measurements of the IMU with the visual structure and performing pre-integration,extracting feature points from the point cloud data of the Li DAR and correcting the motion distortion of the point cloud,analyzing the Cartesian coordinate representation of the GPS,and providing the required initial state for motion estimation through initialization.Afterwards,the system runs two threads of visioninertia odometry and Li DAR odometry in parallel in the local estimation layer to estimate the local motion(relative pose)of the vehicle.Subsequently,in the global optimization layer,the absolute geographic position information of GPS is introduced to eliminate the accumulated drift in the local estimation layer.Finally,when the global optimization is performed,the accuracy of the motion estimation is further improved through the pose graph optimization and the pose is converted to the global coordinate system.In this paper,sufficient public dataset tests and real-world environment tests have been carried out,and detailed performance comparison analysis has been made.First,when comparing the performance with single-sensor algorithms,through 9 sets of KITTI dataset tests,it is proved that the estimation effect of the fusion algorithm is better than that of the single-sensor algorithms.The subsequent test results of Eu Ro C indoor dataset and HKUST outdoor dataset show that the performance of this method is better than the current commonly used fusion algorithms.Finally,the robustness and practicability of the method are evaluated through the campus environment test and actual vehicle test in a certain region of a research institute.The final statistical results show that the average translation error of this method is 0.8045% and the average rotation error is 0.0043deg/m.