基于多模态感知与融合的无人车韧性导航系统

针对野外复杂环境下的无人车自主导航需要,建立了一种基于多源融合定位、语义建图与运动规划的智能导航系统.首先,针对IMU、轮式里程计、视觉SLAM与激光雷达SLAM等测量子系统,设计了误差状态扩展卡尔曼滤波器进行融合定位.其次,基于改进的CNN语义分割网络生成环境的语义图像,与3D激光雷达点云融合,并使用最大概率更新算法构建语义3D地图.接着,在语义和几何信息投影获得可通行性代价的基础上,提出了一种语义动态窗口的局部路径规划方法.最后,将以上感知、定位与规划方法整合成完整的智能导航系统,在城市与野外典型场景的测试中,相对定位误差小于0.4％D,具备一定的韧性导航定位和智能感知规划能力.

Flexible Navigation System for Unmanned Ground Vehicle Based on Multi-modal Sensing and Fusion

An intelligent navigation system based on multi-source fusion positioning, semantic mapping and motion planning is established to meet the needs of autonomous navigation of unmanned ground vehicles in complex field environments. Firstly, the measurement subsystems, such as IMU, wheeled odometer, visual SLAM and lidar SLAM, are fused for localization by error state extended Kalman filter. Secondly, the semantic image of the environment is generated based on the improved CNN semantic segmentation network, which is fused with the 3D lidar point cloud, and the semantic 3D map is constructed using the maximum probability update algorithm. Then, based on the passage cost computed by the projection of semantic and geometric information, a local path planning method based on semantic dynamic window is proposed. Finally, the above sensing, positioning and planning methods are integrated into a complete intelligent navigation system. In the tests of typical urban and wild scenes, the relative positioning error is less than 0.4%D, and it has a certain ability of tough navigation positioning and intelligent perception planning.