Research On Navigation Based On Multi-sensor Fusion SLAM

In recent years,the development of autonomous driving technology has been rapid,and simultaneous localization and mapping(SLAM)technology is a key technology in the field of autonomous driving.Many scholars are working hard to develop this technology.This article studies the defects in loop detection modules in laser SLAM and proposes to improve the loop detection capability of laser SLAM by adding visual image information from a camera.This can avoid or reduce missed detections and false alarms,and construct point cloud maps that are more suitable for autonomous driving requirements.Considering that the system needs to be deployed on the Jetson AGX Xavier platform,which has limited computing resources,this article focuses on the LeGO-LOAM framework,which performs well and is lightweight in laser SLAM,and analyzes its principles and characteristics in detail.Based on the test platform of the Baojun E300 vehicle,the system was tested in two different experimental areas.Subjective and objective analysis showed that LeGOLOAM performed well in test park 1,which was a small area with clear geometric features,but in test park 2,which was a larger area with similar geometric features,the map did not loop back properly,demonstrating that LeGO-LOAM has poor loop detection performance in large areas with similar geometric features.To address the above issues,this article proposes a multi-sensor fusion SLAM system,which adds a visual loop detection module to the LeGO-LOAM framework.By formulating a camera image keyframe selection strategy,selecting camera image keyframes,and extracting ORB feature points from the camera images,an ORB vocabulary bag is created.Visual loop detection is performed using the selected image keyframes and the constructed ORB vocabulary bag,and the result is used to improve the loop detection of Lego-LOAM,thus achieving the goal of building a point cloud map with good global consistency.Through tests on the experimental platform,it was found that the addition of the visual loop module did not significantly affect mapping performance in small areas with clear geometric features,but effectively improved loop detection performance in large areas with similar geometric features,resulting in better global consistency of the constructed point cloud map.Finally,based on the multi-sensor fusion SLAM system constructed in this paper,a point cloud map was built and a Lanelet2 map was drawn.The Autoware.universe navigation framework was deployed using Jetson AGX Orin.Navigation experiments were completed in test park 1 to validate the usability and accuracy of the point cloud map constructed by the multi-sensor fusion SLAM system proposed in this paper from the perspective of whole vehicle navigation.