Direct method, the typical representative of direct method is ICP and NDT series. ICP can directly calculate the position and attitude of laser, and multiple frames can form joint optimization. This scheme is simple and effective, and is often used for multi-path alignment of lidar.
Based on feature matching (LO), the typical representative of this scheme is LOAM and the subsequent improved scheme A-LOAM/F-LOAM. This scheme calculates the pose between frames by finding line, surface features and matching features, and can optimize multiple poses by BA.
Multi-sensor fusion scheme. The typical representatives of this scheme are LIO- mapping, LINS and LIO- Sam. LIO mapping algorithm draws lessons from VINS- Mono's pre-integration and back-end optimization, and the front-end visual odometer is changed to laser odometer.
Based on grid, the representative of this scheme is Google's open source cartography, which has advantages in indoor robot positioning.
Based on panel, the typical representative of this scheme is suma.
Based on semantic information, the typical representatives of this scheme are segmap and suma++.
The production of high-precision maps includes semantic information extraction (usually from vision, but also from laser), one-way LIO(GNSS+IMU+DMI+ lidar/visual odometer) and multi-path alignment. Let's talk about the difficulties of each part first:
Extraction of image semantic information. The semantic information of the image includes lane lines, rod-shaped objects, signs, ground traffic signals and so on. In the high-speed scene, there is less object occlusion, and the accuracy of object detection can reach more than 95%. In urban roads, it is difficult to achieve more than 90% detection because of the occlusion of trees (the occlusion of signs and rod-shaped objects) and vehicles (the occlusion of lane lines and ground traffic signals).
Semantic information extraction of point cloud. Point clouds have high reflectivity for special materials (such as lane lines), and basically dichotomy can solve many problems. For high-speed scenes, lane line wear is not serious, and semantic information is easy to extract. It is difficult to extract more than 90% semantic information because of the serious wear of urban road lanes and the interference of old car lanes.
LIO。 High-precision collection vehicles are generally equipped with lidar, camera, imu, dmi and RTK, and the one-way trajectory can be fused by multiple sensors. For high-speed scenes, the buildings are less shaded, and the RTK signal is better. After construction, the RTK accuracy can be less than 30cm (except for mountain and tunnel scenes), and the automation is difficult, which is related to the scene.
Multi-pass fusion mainly depends on the human eye to distinguish whether the point clouds are aligned. Of course, small-scale data sets can be established for evaluation, and the automation rate is very low.
Generally speaking, it is difficult to realize the automatic production of high-precision maps at present. The main reason is that the scene is complex, there are too many corners, and the absolute accuracy and relative accuracy are difficult to meet the requirements.
To sum up, the direct method of laser SLAM is relatively simple at present, and can be used for multi-path alignment or loop detection of laser odometer. At present, pure LO algorithm is rarely used in industry, and multi-sensor fusion scheme is generally adopted. After all, the lidar is used, which is not bad for a sensor like IMU. Multi-sensor fusion scheme is mainly used to make high-precision maps. In the field of autonomous driving, high-precision maps are generally used as a priori to locate, and they will not locate and map at the same time. Grid-based can be used for mobile robots. In indoor environment, grid assumption is mostly effective. In outdoor scenes, NDT maps are generally used to store maps. Based on semantic information and Ning Bin, people don't know much about it, and they don't use it much in industry.