Localization Based on Semantic Map and Visual Inertial Odometry（ICPR2018）
The perception result establishes a priori map (ICP+GPS), based on the VINS-MONO framework, integrated into GPS, and uses the 2D perception results of each frame to match in the 3D map, and optimizes the pose through the design of semantically constrained reprojection errors.
2D lane line: fit as a curve 
3D lane line:
traffic sign:
triangle sign saves the coordinates of three points,
rectangle sign saves the coordinates of four points,
circle sign saves the center coordinates and radius
Light pole: the starting coordinates of a straight line
Using the GPS global pose as the initial value, determine the search radius, obtain candidate 3D coordinates within the range, reproject the 3D landmark to the current camera coordinates, and reproject the error according to the different definitions of the semantic tag category. After determining the matching relationship, continue to track this For 2D-3D matching, if it appears more than once, it is considered to be an available constraint.
Because lane lines and light poles are easy to mismatch, the characteristics are insufficient, and the initialization is mainly performed by marking 2D-3D matching. The initial matching method is to select the most suitable according to the area. A bunch of flags, and then find the smallest error among the candidates.
Because the matching is relatively sparse, the initialization will be completed after multiple frames. This constraint is used to optimize
Re-projection error definition:
Lane line: the distance from the 3D point to the 2D curve, and the difference in the x direction is used as the distance 
mark. The
circle mark takes the distance from the center of the circle as the error
Lights:
Monocular Localization in Urban Environments using Road Markings
A priori global map, generated by radar, manually labeling landmarks, using only solid and dashed information to indicate. When extracting a single frame, only mark edge lines within 80m in length and 15m in width are used for matching. The optimization part is mainly It consists of three constraints: chamfer matching, opposite pole constraint, and odometer.
Chamfer matching: all candidate 3D lane lines are projected onto the current picture, and the chamfer matching
epipolar constraint between them and the edge line picture is calculated. Constraints: the ones that match the same 3D point before and after the frames are regarded as matched pairs, forming a pair of five degrees of freedom Extreme constraint (monocular scale problem)
Odometer constraint: the amount of change in translation direction
Initialization: first use the value of the odometer and randomly extract candidate point clouds, make the candidate pose with the smallest cost function of the three constraints as the initialization value, and reinitialize it whenever the chamfer matching distance is too large
AVP-SLAM
There are stop lines, speed bumps, and arrows in the semantic prior map
Four cameras surround view+imu+encoder
Each frame has a 3D local map. There are 3D-3D ICP matching between the previous and next frames. There are two constraints: 1. Constraints between adjacent frames to optimize the local odometer, 2. Loop detection with the global map to optimize the global mileage Calculate and calculate the projection error.
Initialization: Mark the entrance of the parking lot on the map, or use the GPS before entering the parking lot as the initial value
. For places that are too abnormal and untextured, use EKF to smooth the trajectory
Road-SLAM : Road marking based SLAM with lane-level accuracy(IV2017)
Consider the error of semantic map, Binocular extraction of perception results, combined with imu and encoder to build a global map, once a single frame is detected, a sub-map and global map are built for 3D-3D matching, and there is a search range based on the initial positioning, which is matched by ICP , Calculate the relative position of the vehicle in the two maps, if a loop is found, the entire map will be optimized.
Semantic signs used: dashed lanes, arrows on the ground, signs, numbers