High-precision map construction method, electronic device, and storage medium

An electronic device, disposed on a vehicle in a field of autonomous driving and comprising:

• at least one processor;
• and
• a memory communicatively connected to the at least one processor,
• wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to execute:
• calculating a pose of a camera at each position point according to a pre-acquired video acquired by devices that have lower cost and are more feasible for large-scale deployment compared to laser devices;
• calculating an absolute depth of each keypoint in the pre-acquired video according to the pose of the camera at each position point;
• according to the absolute depth of each keypoint in the pre-acquired video, constructing a corresponding three-dimensional point cloud of each pixel point in the pre-acquired video;
• and
• according to the corresponding three-dimensional point cloud of each pixel point in the pre-acquired video, constructing a high-precision map corresponding to the pre-acquired video to provide reliable road information for the autonomous driving, and ensure safety of the autonomous driving;
• wherein the at least one processor is configured to:
• extract two images from the pre-acquired video as a current image pair, wherein the current image pair comprises a first current image and a second current image;
• extract at least one keypoint from the first current image and at least one keypoint from the second current image, and match the at least one keypoint in the first current image with the at least one keypoint in the second current image one by one to obtain at least one successfully-matched keypoint;
• calculate an absolute depth of each of the at least one successfully-matched keypoint in the first current image according to a pose of the camera at a position point at which the first current image is shot, and calculate an absolute depth of each successfully-matched keypoint in the second current image according to a pose of the camera at a position point at which the second current image is shot;
• and
• wherein the at least one processor is configured to:
• construct a pixel coordinate of the first current image and a normalized plane coordinate of pixel points in the first current image according to the pose of the camera at the position point at which the first current image is shot, and construct a pixel coordinate of the second current image and a normalized plane coordinate of pixel points in the second current image according to the pose of the camera at the position point at which the second current image is shot;
• and
• calculate the absolute depth of each of the at least one successfully-matched keypoint in the first current image according to the pixel coordinate of the first current image and the normalized plane coordinate of the pixel points in the first current image, and calculate the absolute depth of each of the at least one successfully-matched keypoint in the second current image according to the pixel coordinate of the second current image and the normalized plane coordinate of the pixel points in the second current image;
• wherein the at least one processor is configured to:
• filter pixel points with inconsistent absolute depths in the pre-acquired video through comparing a consistency relationship between absolute depths of diffused pixel points and an absolute depth of each successfully-matched keypoint in the pre-acquired video;
• and
• wherein the at least one processor is configured to:
• obtain three-dimensional point clouds of a road scene that comprise semantic information from each image in the pre-acquired video;
• obtain a point cloud cluster belonging to a same category by combining the three-dimensional point clouds of the road scene;
• perform a vectorization operation on the obtained point cloud cluster to obtain a vectorization result, wherein the vectorization result is classified into a three-dimensional type, a planar type, and a linear type;
• for the three-dimensional type, the vectorization operation is performed using three-dimensional bounding boxes of outer boundaries of the point cloud cluster;
• and for the planar type and the linear type, the vectorization operation is performed by defining the three-dimensional bounding boxes of the outer boundaries of the point cloud cluster and compressing the three-dimensional bounding boxes to form a two-dimensional planar vectorization result.

A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are configured to cause a computer to execute:

• calculating a pose of a camera at each position point according to a pre-acquired video acquired by devices that have lower cost and are more feasible for large-scale deployment compared to laser devices;
• calculating an absolute depth of each keypoint in the pre-acquired video according to the pose of the camera at each position point;
• according to the absolute depth of each keypoint in the pre-acquired video, constructing a corresponding three-dimensional point cloud of each pixel point in the pre-acquired video;
• and
• according to the corresponding three-dimensional point cloud of each pixel point in the pre-acquired video, constructing a high-precision map corresponding to the pre-acquired video to provide reliable road information for the autonomous driving, and ensure safety of the autonomous driving;
• wherein the computer calculates the absolute depth of each keypoint in the pre-acquired video according to the pose of the camera at each position point by:
• extracting two images from the pre-acquired video as a current image pair, wherein the current image pair comprises a first current image and a second current image;
• extracting at least one keypoint from the first current image and at least one keypoint from the second current image, and matching the at least one keypoint in the first current image with the at least one keypoint in the second current image one by one to obtain at least one successfully-matched keypoint;
• and
• calculating an absolute depth of each of the at least one successfully-matched keypoint in the first current image according to a pose of the camera at a position point at which the first current image is shot, and calculating an absolute depth of each successfully-matched keypoint in the second current image according to a pose of the camera at a position point at which the second current image is shot;
• and
• wherein the computer calculates the absolute depth of each of the at least one successfully-matched keypoint in the first current image and the absolute depth of each of the at least one successfully-matched keypoint in the second current image by:
• constructing a pixel coordinate of the first current image and a normalized plane coordinate of pixel points in the first current image according to the pose of the camera at the position point at which the first current image is shot, and constructing a pixel coordinate of the second current image and a normalized plane coordinate of pixel points in the second current image according to the pose of the camera at the position point at which the second current image is shot;
• and
• calculating the absolute depth of each of the at least one successfully-matched keypoint in the first current image according to the pixel coordinate of the first current image and the normalized plane coordinate of the pixel points in the first current image, and calculating the absolute depth of each of the at least one successfully-matched keypoint in the second current image according to the pixel coordinate of the second current image and the normalized plane coordinate of the pixel points in the second current image;
• wherein the computer instructions are configured to cause the computer to further execute:
• filtering pixel points with inconsistent absolute depths in the pre-acquired video through comparing a consistency relationship between absolute depths of diffused pixel points and an absolute depth of each successfully-matched keypoint in the pre-acquired video;
• and
• wherein the computer constructs the high-precision map corresponding to the pre-acquired video according to the corresponding three-dimensional point cloud of each pixel point in the pre-acquired video by:
• obtaining three-dimensional point clouds of a road scene that comprise semantic information from each image in the pre-acquired video;
• obtaining a point cloud cluster belonging to a same category by combining the three-dimensional point clouds of the road scene;
• performing a vectorization operation on the obtained point cloud cluster to obtain a vectorization result, wherein the vectorization result is classified into a three-dimensional type, a planar type, and a linear type;
• for the three-dimensional type, the vectorization operation is performed using three-dimensional bounding boxes of outer boundaries of the point cloud cluster;
• and for the planar type and the linear type, the vectorization operation is performed by defining the three-dimensional bounding boxes of the outer boundaries of the point cloud cluster and compressing the three-dimensional bounding boxes to form a two-dimensional planar vectorization result.