CN-121977590-A - Complex intersection map making method adopting multi-sensor fusion scheme

CN121977590ACN 121977590 ACN121977590 ACN 121977590ACN-121977590-A

Abstract

The invention discloses a complex intersection map making method of a multi-sensor fusion scheme, which belongs to the technical field of intelligent driving and comprises the steps of preprocessing single-pass data to generate an odometer track, aligning the odometer track with a GPS positioning track to obtain a single-pass track, screening point cloud frame pairs with center point distances smaller than a preset distance threshold value from the single-pass track to form a potential point cloud matching pair set of cross-pass matching, carrying out two-stage registration on each point cloud frame pair in the set based on a registration network and an algorithm, solving the relative pose of each point cloud frame, constructing a map model based on the relative poses of all the point cloud frames, introducing the same-pass odometer track constraint and the cross-pass relative pose constraint to carry out joint optimization, and fusing each pass point cloud based on the optimized pose to generate an intersection point cloud map. The invention can realize the construction of the complex intersection point cloud map with high precision, strong robustness, automation and low cost, and is suitable for application scenes with high requirements on map freshness and precision, such as intelligent traffic, automatic driving and the like.

Inventors

ZHAO YUTING
LI XIAOCHUAN
WANG XINGBO
JIANG LINGXIN

Assignees

北京大学

Dates

Publication Date: 20260505
Application Date: 20260213

Claims (10)

1. The complex intersection map making method of the multi-sensor fusion scheme is characterized by comprising the following steps of: Preprocessing single-pass data of each acquisition vehicle passing through a complex intersection to generate an odometer track, and aligning the odometer track with a GPS positioning track to a unified coordinate system to obtain an aligned single-pass track; Based on the spatial proximity, automatically screening point cloud frame pairs with center point distances smaller than a preset distance threshold value from different single-pass tracks to form a potential point cloud matching pair set of cross-pass matching; for each point cloud frame pair in the potential point cloud matching pair set, firstly adopting a registration network to perform coarse registration, and then adopting a registration algorithm to perform fine registration so as to solve the relative pose of each point cloud frame; Constructing a graph model based on the relative poses of all the point cloud frames, introducing the trajectory constraint of the same-pass odometer and the relative pose constraint of the cross-pass, and carrying out joint optimization on all the poses by minimizing the integral constraint residual error; and fusing the point clouds of each single-pass data by using the optimized pose to generate an intersection point cloud map.
2. The method for making a complex intersection map according to the multi-sensor fusion scheme of claim 1, wherein the step of automatically screening point cloud frame pairs with center point distances smaller than a preset distance threshold from different single-pass tracks based on spatial proximity to form a set of potential point cloud matching pairs for cross-pass matching comprises the steps of: and placing the central point of each frame of point cloud in the aligned single-pass track under a unified coordinate system, calculating Euclidean distance between the central points of any two frames of point clouds from different single-passes, taking the point cloud frame pairs with the distance smaller than a preset distance threshold as candidate matching pairs, and adding a potential point cloud matching pair set.
3. The method for mapping complex intersections of a multi-sensor fusion scheme according to claim 1, wherein the coarse registration using a registration network comprises: Dividing each point cloud frame pair in the potential point cloud matching pair set into a source point cloud and a target point cloud to be registered, and inputting a registration network constructed based on a deep learning network; In a registration network, firstly, carrying out multi-stage downsampling on an input point cloud by utilizing a downsampling layer to form a hierarchical structure comprising the input point cloud, dense point clouds and rough point clouds, then utilizing a Transformer layer to extract local features of each dense point cloud and concatenate the local features into local features of the rough point cloud, utilizing Geometric Transformer layers to extract geometric features based on the local features of the rough point cloud, then utilizing an attention mechanism to fuse the geometric features of the rough point cloud to respectively obtain source point cloud features and target point cloud features, then calculating a Gaussian correlation matrix between the source point cloud features and the target point cloud features, selecting Top K matched point pairs based on matrix element values, finally calculating a local transformation matrix based on the dense point clouds corresponding to each matched point pair, and selecting an optimal local transformation matrix as a rough registration output result from the source point cloud to the target point cloud by utilizing a voting mechanism.
4. A complex intersection mapping method according to claim 1 or 3, wherein the fine registration using a registration algorithm comprises: Taking a local transformation matrix obtained by rough registration as initial estimation, and adopting a random sampling consistency algorithm to remove mismatching from the point cloud corresponding relation to obtain an internal point set; Based on the interior point set, iterative optimization is carried out by adopting an iterative nearest point algorithm, and finally an optimal relative pose transformation matrix is solved and used as a fine registration output result from a source point cloud to a target point cloud.
5. The method for making the complex intersection map by the multi-sensor fusion scheme according to claim 1, wherein the constructing a map model based on the relative pose of all the point cloud frames and introducing the same-trip odometer track constraint and the cross-trip relative pose constraint comprises the following steps: the relative pose of the point cloud frame acquired by each acquisition vehicle at each moment is defined as each node in the graph model, and the pose constraint among the nodes is defined as the edge in the graph model; Wherein, edge connection is constructed between the continuous frame nodes belonging to the same single pass through the same-pass odometer track constraint, edge connections are constructed between the registered frame nodes from different single passes through cross-pass relative pose constraints.
6. The complex intersection mapping method of a multi-sensor fusion scheme according to claim 1 or 5, wherein the joint optimization of all poses by minimizing an overall constraint residual comprises: defining an optimization objective function as follows: , Wherein, the Representing the pose of the multi-pass trajectory, being the optimization variable of the objective function, Represent the first In the middle of the pass And Two frames of the same-pass odometer track constraint, Represent the first First pass Frame and the first First pass The cross-pass relative pose constraint of the frame, And Respectively representing the track constraint weight of the same-trip odometer and the relative pose constraint weight of the crossing trip, And Frame index sets representing the same pass and the cross pass, respectively, are superscript Representing a transpose; , Wherein, the Represent the first In the middle of the pass And The relative pose measurement of two frames, Represent the first In the middle of the pass And The relative pose estimate of the two frames, Representing a unified coordinate system; , Wherein, the Represent the first First pass Frame and the first First pass The relative pose measurements of the frames, Represent the first First pass Frame and the first First pass Estimating the relative pose of the frame; optimizing variables by iterative adjustment To minimize the objective function Thereby obtaining the globally consistent optimal pose estimation.
7. The method for making the complex intersection map according to the multi-sensor fusion scheme of claim 1, wherein the fusing the point clouds of each single pass of data by using the optimized pose to generate the intersection point cloud map comprises: transforming the original point cloud of each single frame into a unified world coordinate system by using the optimized pose, superposing, performing point cloud filtering and voxel downsampling on the whole point cloud after superposition, and generating a final three-dimensional point cloud map of the complex intersection.
8. The complex intersection map making device adopting the multi-sensor fusion scheme is realized by the complex intersection map making method adopting the multi-sensor fusion scheme as claimed in any one of claims 1-7, and is characterized by comprising a track preliminary alignment module, a cross-trip point cloud association module, a relative pose solving module, a global pose optimizing module and a point cloud map generating module; The track preliminary alignment module is used for preprocessing single-pass data of each acquisition vehicle passing through a complex intersection, generating an odometer track, and aligning the odometer track with the GPS positioning track to a unified coordinate system to obtain an aligned single-pass track; the cross-trip point cloud association module is used for automatically screening point cloud frame pairs with center point distances smaller than a preset distance threshold value from different single-trip tracks based on spatial proximity to form a cross-trip matching potential point cloud matching pair set; The relative pose solving module is used for carrying out coarse registration on each point cloud frame pair in the potential point cloud matching pair set by adopting a registration network, and then carrying out fine registration by adopting a registration algorithm so as to solve the relative pose of each point cloud frame; the global pose optimization module is used for constructing a graph model based on the relative poses of all point cloud frames, introducing the trajectory constraint of the same-pass odometer and the cross-pass relative pose constraint, and carrying out joint optimization on all poses by minimizing the overall constraint residual error; The point cloud map generation module is used for fusing the point clouds of each single-pass data by utilizing the optimized pose to generate an intersection point cloud map.
9. An electronic device comprising a memory for storing a computer program and one or more processors, wherein the processor is configured to implement the complex intersection mapping method of the multisensor fusion scheme of any one of claims 1 to 7 when the computer program is executed.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that the complex intersection mapping method of the multisensor fusion scheme of any one of claims 1 to 7 is implemented when the computer program is executed by a computer.

Description

Complex intersection map making method adopting multi-sensor fusion scheme Technical Field The invention belongs to the technical field of map making, and particularly relates to a complex intersection map making method of a multi-sensor fusion scheme. Background Aiming at the requirements of environment accurate perception and decision making which are necessary for realizing high-level intelligent driving, the construction and maintenance of a high-precision map becomes a core support technology, and is a data base stone for realizing accurate positioning, reliable perception and prospective planning of an automatic driving system. At present, the high-precision map making in the prior art mainly comprises two major schemes, namely a crowdsourcing scheme based on vision and a crowdsourcing scheme based on a multi-sensor fusion and update strategy, which have achieved certain effects in open scenes. The crowdsourcing scheme based on vision is to take a vehicle-mounted camera as a main sensor, acquire an image sequence acquired by multiple vehicles in a crowdsourcing mode, and construct a map by adopting a computer vision technology. The patent document with publication number CN110060343a proposes a map construction method, the process mainly comprising image acquisition, local map construction, similar image detection, relative pose estimation, and global map construction. The method is based on the extraction and matching of visual features, and aims to realize the construction and updating of a map in a large range at a low cost. In order to make up for the limitation of a single sensor, the prior art fuses a GNSS, an inertial navigation system and a visual sensor and introduces an intelligent updating strategy. A close-coupled approach is proposed by the patent document with publication number CN110412635A, which uses a priori a visual feature point cloud map as an environmental beacon to provide positioning constraints through visual matching when GNSS signals are occluded. Aiming at a specific scene, a dynamic crowdsourcing updating strategy based on contribution degree discrimination is provided in the patent document with the publication number of CN116089445A, so that efficient and on-demand incremental updating of a map is realized, and the interference to continuous operation of a vehicle is reduced. The core of the scheme is that the balance between the positioning precision and the updating efficiency is sought through multi-source information complementation and flow optimization. Although the scheme has certain feasibility in general road scenes, the following significant problems still exist in complex intersection scenes such as viaducts, multilayer overpasses, large roundabout and the like: (1) The visual matching is invalid under the repeated texture, namely the repeated texture is easy to exist under the scenes such as an overhead bridge and the like, such as a continuously consistent isolation belt, visual characteristics cannot be obviously distinguished, mismatching is easy to occur, a local map splicing error is caused, and ghost images and structural dislocation are finally generated in a global map. (2) The multi-sensor fusion is restricted by signal degradation, and GNSS signals at complex intersections are easy to be blocked and interfered by multipath effects, so that positioning drift is generated. If the point cloud map is directly aligned and overlapped, a system error is introduced, so that dislocation and ghost images appear between map layers acquired in different periods. (3) The updating strategy is not suitable for the complexity and real-time requirements of the crossing, the existing incremental updating mechanism is difficult to accurately judge the real change and the perceived noise in the scene of complex crossing structure and dense traffic flow, and if the updating time is incorrect or the updating content is inaccurate, the navigation decision error is easily caused. In summary, the existing crowdsourcing map construction and update scheme still has obvious defects in aspects of visual feature matching, multi-sensor fusion positioning precision, intelligent update strategy under a high dynamic environment and the like in a complex intersection scene, and the reliable application of a high-precision map in an automatic driving system is restricted. Therefore, a map making method that can effectively fuse multi-source sensing data, adapt to the structural characteristics of complex intersections, and has strong robustness and real-time updating capability is needed. Disclosure of Invention In view of the above, the invention aims to provide a complex intersection map making method of a multi-sensor fusion scheme, which utilizes laser radar data collected by a common passenger train to construct a cross-trip intelligent association by fusing single-trip data collected by multiple vehicles when passing through the same intersection at different time periods, const