CN-121994203-A - Semantic map construction method and system based on multiple input information

CN121994203ACN 121994203 ACN121994203 ACN 121994203ACN-121994203-A

Abstract

The invention provides a semantic map construction method and a semantic map construction system based on multiple input information, wherein the method comprises the steps of inputting data, acquiring a point cloud map, GNSS data with a time stamp and image information with the time stamp; extracting and vectorizing semantic elements, extracting the semantic elements from the point cloud map by using a deep learning model, and storing the semantic elements as vectorized elements, wherein the vectorized elements comprise point elements representing specific points on the map, line elements which are formed by a series of continuous point elements and are used for representing linear characteristics, polygonal elements which are formed by a series of point elements connected end to end and are used for representing polygonal elements with definite boundary areas, manually correcting the extracted semantic elements, and outputting the corrected semantic elements into a semantic map in a standard format. The invention has the advantages of high detection efficiency, high accuracy and the like, and can effectively improve the efficiency and the accuracy of map construction.

Inventors

LI CHENG
Li Yuanzhengyu
YUAN XIWEN
PAN WENBO
HUANG WENYU
CHEN ZHIWEI

Assignees

中车株洲电力机车研究所有限公司

Dates

Publication Date: 20260508
Application Date: 20241101

Claims (10)

1. The semantic map construction method based on the multi-input information is characterized by comprising the following steps of: the method comprises the steps of data input, obtaining a point cloud map, GNSS data with a time stamp and image information with a time stamp, wherein the point cloud map is used for constructing a reference base map of a semantic map, the GNSS data with the time stamp is used for constructing time sequence track information of a vehicle in a travelable area in a line, and the image information with the time stamp is used for image reference of semantic element extraction; Extracting and vectorizing semantic elements, extracting the semantic elements from the point cloud map by using a deep learning model, and storing the semantic elements as vectorized elements, wherein the vectorized elements comprise point elements representing specific points on the map, line elements which are formed by a series of continuous point elements and are used for representing linear characteristics, and polygonal elements which are formed by a series of point elements which are connected end to end and are used for representing a definite boundary area, and manually correcting the extracted semantic elements and outputting the corrected semantic elements into a semantic map in a standard format.
2. The semantic map construction method based on multiple input information according to claim 1, wherein the point elements include waypoint vector elements, the extraction of the waypoint vector elements is to perform downsampling processing on time sequence track information of a vehicle in a travelable area in a line according to a certain distance interval to obtain a plurality of waypoint information, extract geographic information, distance information and speed and attribute information of each waypoint information, and store the extracted waypoint information as a vector format to form the waypoint vector elements.
3. The semantic map construction method based on multiple input information according to claim 1, wherein the point elements further comprise signal light vector elements, the signal light vector elements are extracted by extracting signal light positions of image information acquired by vehicles at different time points, converting signal light position information in the images into absolute position information in the real world by using GNSS track data of the vehicles and calibration parameters between a camera and a GNSS acquisition device, and storing the position information as vectorized point elements.
4. The semantic map construction method based on multiple input information according to claim 3, wherein after absolute position information of the signal lamp is obtained, final position of the signal lamp is obtained by performing weighted average processing on signal lamp position information collected under a plurality of different vehicle postures, wherein the closer the distance between the vehicle and the signal lamp is, the greater the weight of the signal lamp position information is, and if the distance between the vehicle and the signal lamp exceeds a threshold value, the signal lamp position information is abandoned at the moment.
5. The semantic map construction method based on multiple input information according to claim 1, wherein the line elements include lane line vector elements, and the process of extracting the lane line vector elements includes: Dividing the point cloud base map into a plurality of grids, wherein each grid contains a specified amount of point cloud data; mapping the point cloud intensity information in the segmented point cloud base map into a two-dimensional gray image, so as to obtain a gray image representing the point cloud intensity distribution; processing the gray level image by using a deep learning method to extract lane line characteristics, and separating out point cloud data belonging to lane lines in a point cloud base map according to the lane line characteristics to form lane line point clouds; And combining the lane line point cloud data in different grids to form a complete lane line point cloud set, fitting the combined lane line point cloud data into a continuous curve by using a curve fitting technology, and storing the continuous curve as a discretized vector line element.
6. The multi-input information based semantic map construction method according to claim 1, wherein the polygonal elements include barrier vector elements, and the extraction process of the barrier vector elements includes: Dividing the point cloud base map into a plurality of grids, wherein each grid contains a specified amount of point cloud data; And processing the segmented point cloud bottom graph by using a deep learning method to obtain an obstacle point cloud, determining the category of the obstacle, and storing the obstacle as a vectorized polygonal element.
7. The multi-input information based semantic map construction method according to claim 1, wherein the polygonal elements further comprise a station vector element, and the station vector element extraction process comprises: Dividing the point cloud base map into a plurality of grids, wherein each grid contains a specified amount of point cloud data; And processing the point cloud data in each grid by using a deep learning method to obtain a platform area, and storing the platform area as vectorized polygonal elements, wherein the polygonal elements comprise specific boundary position information of the platform and building types to which the platform belongs.
8. The semantic map construction method based on multiple input information according to any one of claims 1 to 7, wherein the point cloud map is constructed by laser radar and combined inertial navigation, and radar reflection intensity information and point cloud intensity information are acquired by adopting millimeter wave radar.
9. The semantic map construction method based on multi-input information according to claim 1, wherein the manual correction is to import the vector elements into a visual map editing platform and perform correction processing on the extracted semantic elements.
10. A multiple-input information based semantic map system comprising a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to perform the multiple-input information based semantic map construction method of any one of claims 1-9.

Description

Semantic map construction method and system based on multiple input information Technical Field The invention relates to the technical field of semantic map and intelligent driving, in particular to a semantic map construction method and system based on multi-input information. Background In the fields of intelligent driving, intelligent rail, mine cards, rail transit and the like, the semantic map is taken as an important environment information carrier, and plays an important role in the functions of sensing, decision planning, control, prediction, positioning and the like of vehicles. A variety of semantic map system construction methods have been developed in the industry, most of which employ a multi-sensor fusion technology path. Such sensors include, but are not limited to, integrated inertial navigation Systems (IMUs), liDAR (LiDAR), millimeter-wave radar, vision sensors, and the like. By fusing the data of these sensors, a semantic map containing rich environmental information can be constructed. In the prior ART, for example, in chinese patent application CN118115963a, which is a method for constructing an online semantic vector map based on a navigation map and visual information, the method can meet the requirement of an automatic driving vehicle on environmental understanding to a certain extent, but because of the manner of constructing a local map in real time based on deep learning, the requirement on computing resources is higher, and in specific application scenarios such as an intelligent track bus system (ART), a mining truck, and track traffic, the problem of excessively redundant map elements may be caused due to high environmental complexity. On the other hand, the Chinese patent application No. CN118031983A discloses an automatic driving fusion positioning method and system, the method realizes accurate positioning of vehicles based on IMU, laser point cloud data and vector map, and the method can improve the positioning accuracy, but has the main focus on positioning rather than comprehensive semantic information extraction and processing for the above-mentioned application scenes. In addition, in the chinese patent application CN117893853a, the method for model training, the method for map generation, the device and the storage medium, the method can fuse the local vectorized map to generate the global vectorized map, but in specific industrial application scenarios, because these scenarios often have the characteristics of fixed route and strong repeatability, the systematic requirements for the map may not be fully satisfied. Disclosure of Invention Aiming at the problems in the prior art, the invention provides a semantic map construction method and a semantic map construction system based on multi-input information, which have high detection efficiency and high accuracy. In order to solve the technical problems, the invention adopts the following technical scheme: a semantic map construction method based on multi-input information comprises the following steps: the method comprises the steps of data input, obtaining a point cloud map, GNSS data with a time stamp and image information with a time stamp, wherein the point cloud map is used for constructing a reference base map of a semantic map, the GNSS data with the time stamp is used for constructing time sequence track information of a vehicle in a travelable area in a line, and the image information with the time stamp is used for image reference of semantic element extraction; Extracting and vectorizing semantic elements, extracting the semantic elements from the point cloud map by using a deep learning model, and storing the semantic elements as vectorized elements, wherein the vectorized elements comprise point elements representing specific points on the map, line elements which are formed by a series of continuous point elements and are used for representing linear characteristics, polygonal elements which are formed by a series of point elements which are connected end to end and are used for representing a polygonal element with a definite boundary area, manually correcting the extracted semantic elements, and outputting the corrected semantic elements into a semantic map in a standard format. The method is further improved in that the point elements comprise road point vector elements, the road points are obtained by carrying out downsampling on time sequence track information of the vehicle in a travelable area in a line according to a certain distance interval to obtain a plurality of road point information, geographic information, distance information and speed and attribute information of each road point information are extracted, and the extracted road point information is stored in a vector format to form the road point vector elements. The method is further improved in that the point elements further comprise signal lamp vector elements, the signal lamp vector elements are extracted, the signal lamp positions