地图辅助的大范围视觉SLAM集束调整理论与方法研究

高精度定位信息至关重要。视觉技术支持车道识别、位姿测量、环境感知，是实现此目标的重要手段。然而，在GNSS长时间难以独立定位时，大范围V-SLAM与V-SLAM/IMU组合系统存在因闭环缺失难以利用集束调整消除累积误差的痛点。.本项目利用V-SLAM、GNSS、数字地图的互补特性，突破建筑物概略模型实时构建方法，研究建筑物模型辅助的SLAM/GNSS紧组合定位算法，提高GNSS无法独立定位时的绝对定位精度；研究车辆异常驾驶模式识别方法，攻克SLAM/路网松组合的多假设多路校正车道级地图匹配算法，提高城市峡谷区的定位可信度；探索地图匹配辅助的视觉SLAM大范围集束调整方法，在线校正V-SLAM，确保GNSS遮挡区的高精度持久定位。.项目预期建立地图辅助的大范围视觉SLAM集束调整理论，开发视觉/GNSS/地图组合定位系统1套，用于城市峡谷区定位时，可在20分钟或更长时间内保持车道级定位精度。

It is very difficult to over-estimate the importance of high-precision positioning information. Computer vision technology, being able to support lane recognition, position and attitude determination, and environment reconstruction concurrently, is thought to be a very important approach to provide such kind of information for vehicle drivers. However, when a vehicle is running in city canyons and the GNSS receiver is not able to provide positioning functionality independently, the pain point for a Visual SLAM system is that it is very difficult to run bundle adjustment to remove accumulated error since there generally is no closure any more.. To overcome the problem, by utilizing the mutual complementary relationships among V-SLAM, GNSS, and digital maps, this project is dedicated on the following researches: through overcoming the method of real-time reconstruction of the building coarse model, by researching the tight integration positioning algorithm of SLAM and GNSS aided by the 3D coarse building models, we expect to improve the absolute positioning ability of the positioning system when the GNSS receiver is not able to provide positioning information; by studying the recognition method of unusual driving styles and conquering road network aided multi-hypothesis-and-multi-correction lane-level-precision map-matching algorithm with SLAM and road network being loose-coupled, we expect to improve the reliability of the positioning results when the vehicle is running in city canyons; by exploring the large-scale bundle adjustment theory of the large-scale SLAM system aided by map-matching, we expect to online correct V-SLAM system and remove the accumulated errors when it is difficult for the GNSS receiver to provide positioning performance, consequently to provide sustainable positioning information where GNSS is heavily blocked.. Through the aforementioned study, our objective is to a new bundle adjustment theory of road network aided large-scale SLAM systems for vehicle positioning in city canyons, and at the same time develop a Camera/GNSS/digital-map integrated low-cost and easy-to-deploy positioning solution, which will keep lane-level-precision positioning performance for more than 20 minutes when the vehicle is running in dense road networks of city canyons.