A Factor Graph Optimization SLAM Mapping Method for Autonomous Vehicles Considering Dynamic Target Motion

In high-level autonomous driving, high-precision map construction is crucial, and map construction based on laser SLAM is one of the mainstream methods. Existing laser SLAM technology usually assumes that the environment is static, ignores dynamic targets, or removes their influence on map construction by directly removing dynamic targets. This paper proposed a factor graph optimization SLAM mapping method for autonomous vehicles considering dynamic target motion, aiming to improve the accuracy and reliability of map construction. First, the kinematic model of the autonomous vehicle was established, and the inertial measurement unit (IMU) and laser radar (LiDAR) were jointly calibrated. The laser point cloud distortion was corrected using the IMU and kinematic model. Then, a semantic spatiotemporal consistency method for dynamic target detection was proposed, and dynamic targets were effectively detected from the laser point cloud through a fully convolutional neural network (FCN), and the estimation of the motion pose of the dynamic target was optimized by combining the improved unscented Kalman filter (UKF). Finally, based on the motion estimation of the dynamic target, the laser point cloud was divided into static and dynamic parts, and the static part was directly registered, while the dynamic part was registered by introducing motion pose compensation. A multi-source asynchronous factor model was constructed through semantic segmentation results, IMU and global positioning system (GPS) data, and a graph optimization method was used for global optimization, which significantly improved the accuracy of map construction and eliminated the negative impact of dynamic targets on map construction.