CN-121979193-A - Autonomous exploration method and system for building space

CN121979193ACN 121979193 ACN121979193 ACN 121979193ACN-121979193-A

Abstract

The embodiment of the disclosure relates to the technical field of unmanned platform autonomous exploration, and provides a building space autonomous exploration method and system, wherein the method comprises the steps of utilizing image point cloud data of a building space to carry out three-dimensional reconstruction to obtain a three-dimensional grid map comprising a plurality of voxels; the image point cloud data comprises visual images and radar point cloud data, a corresponding boundary detection result is obtained by conducting boundary detection on the three-dimensional grid map, door and window detection and room modeling are conducted through the image point cloud data, corresponding door and window semantic information is obtained, a target viewpoint to be executed is generated based on the boundary detection result, an optimal target viewpoint is selected from the target viewpoints to be executed through the door and window semantic information, exploration path planning is conducted based on the optimal target viewpoint, a corresponding exploration track is obtained, and an unmanned platform is enabled to explore a building space according to the exploration track. The embodiment of the disclosure effectively solves the problem that the prior method can search incompletely or backtracking phenomenon occurs so as to greatly reduce the search efficiency.

Inventors

WANG RONG
MENG FANLE
LI JINBO
WU JINLIANG

Assignees

中国电子科技集团公司信息科学研究院

Dates

Publication Date: 20260505
Application Date: 20251217

Claims (10)

1. The autonomous exploration method for the building space is characterized by comprising the following steps of: Performing three-dimensional reconstruction by utilizing image point cloud data of a building space to obtain a three-dimensional grid map comprising a plurality of voxels, wherein the image point cloud data comprises a visual image and radar point cloud data; performing boundary detection on the three-dimensional grid map to obtain a corresponding boundary detection result; Performing door and window detection and room modeling by using the image point cloud data to obtain corresponding door and window semantic information; Generating a target viewpoint to be executed based on the boundary detection result; selecting an optimal target viewpoint from target viewpoints to be executed by utilizing the door and window semantic information; And planning an exploration path based on the optimal target viewpoint to obtain a corresponding exploration track, so that the unmanned platform explores the building space according to the exploration track.
2. The method for autonomous exploration of a building space according to claim 1, wherein said three-dimensional reconstruction using image point cloud data of the building space, to obtain a three-dimensional grid map comprising a plurality of voxels, comprises: Based on the image point cloud data, the three-dimensional grid map represented by the SLAM algorithm construct is used, and the three-dimensional grid map maintains a fixed resolution, so that each voxel in the three-dimensional grid map is ensured to have a fixed side length.
3. The method for autonomous exploration of building space according to claim 1, wherein said performing boundary detection on said three-dimensional grid map to obtain a corresponding boundary detection result comprises: Partitioning and clustering voxels in the three-dimensional grid map according to a certain distance constraint, and marking the category of the voxels corresponding to each cluster as occupied; Marking the category of the voxels on the path between the currently unknown voxels of the unmanned platform and the voxels with the category of "occupied" as "idle"; marking the categories of the rest voxels except the voxels with the category of occupied and the voxels with the category of idle in the three-dimensional grid map as unknown; Taking boundary voxels of the voxels with the category of occupied and unknown as well as boundary voxels of the voxels with the category of idle and unknown as boundary points; And when the boundary area formed by the plurality of boundary points is smaller than or equal to the size of the unmanned platform, removing all the boundary points in the boundary area.
4. The method for autonomous exploration of building space according to claim 1, wherein said performing door and window detection and room modeling using said image point cloud data to obtain corresponding door and window semantic information comprises: parameterizing each room in the building space by using a predefined structure body, and creating a room list, wherein the structure body comprises a door and window size position, wall surface parameters and structural integrity indication; Performing door and window detection by using the image point cloud data, determining the size and the position of a detected door and window by associating the radar point cloud data, if the currently detected door and window can allow the unmanned platform to pass through and the position of the door and window is different from the position of the door and window of any room in the room list, creating a new room in the room list, assigning the size and the position of the currently detected door and window to the door and window size position in the structure of the new room, and simultaneously indicating the structural integrity in the structure of the new room as false so as to indicate that the new room is not completely explored; when detecting doors and windows, extracting a significant plane by using a RANSAC algorithm according to the radar point cloud data, wherein the significant plane is a plane with the point cloud quantity exceeding a certain threshold value; according to the relation between the positions of radar points on the remarkable plane and normal vectors, each room in the room list is associated with a corresponding wall surface; If all the walls of a room in the room list can form a closed cuboid, and no detected boundary area exists inside the cuboid, the structural integrity indication in the structure of the room is made to be true, so that the room is completely explored.
5. The method for autonomous exploration of a building space according to claim 4, wherein said associating each room in said room list with a corresponding wall surface according to a relationship between a position of a radar point located on said salient plane and a normal vector, comprises: If the normal vectors of the two significant planes are perpendicular to each other and intersect in a certain straight line, then the two significant planes are considered to be two walls of a certain room.
6. The method according to claim 1, wherein generating the target viewpoint to be executed based on the boundary detection result includes: And generating a target viewpoint to be executed by adopting a non-shielding sphere generation method based on the boundary detection result.
7. The method for autonomous exploration of building space according to claim 1, wherein said selecting an optimal target viewpoint from target viewpoints to be executed by using said door and window semantic information comprises: respectively determining an evaluation function value of each target viewpoint; Based on the door and window semantic information, the target viewpoint with the largest evaluation function value in the room is preferentially selected as the optimal target viewpoint; if a certain room is not fully explored, continuously selecting the target viewpoint in the room as the optimal target viewpoint until the room is fully explored.
8. The method of autonomous exploration of a building space according to claim 7, wherein said determining an evaluation function value for each of said target viewpoints, respectively, comprises: Determining an evaluation function value of the target viewpoint according to the following formula: ; Wherein, the An evaluation function value representing the target viewpoint, Indicating the amount of information newly added by this search, Representing a distance cost to reach the target viewpoint from the current location where the unmanned platform is located, Representation of Is used for the weight of the (c), Representation of Is a weight of (2).
9. The method for autonomous exploration of building space according to claim 1, wherein said planning exploration paths based on said optimal target viewpoint, obtaining corresponding exploration trajectories, comprises: And based on the optimal target viewpoint, finding out a path which does not collide with the obstacle as the exploration track.
10. A building space autonomous exploration system, the building space autonomous exploration system comprising: The system comprises a voxel map building module, a three-dimensional grid map generation module and a three-dimensional map generation module, wherein the voxel map building module is used for carrying out three-dimensional reconstruction by utilizing image point cloud data of a building space to obtain a three-dimensional grid map containing a plurality of voxels; the boundary detection module is used for carrying out boundary detection on the three-dimensional grid map to obtain a corresponding boundary detection result; the door and window detection and room modeling module is used for carrying out door and window detection and room modeling by utilizing the image point cloud data to obtain corresponding door and window semantic information; the viewpoint generation module is used for generating a target viewpoint to be executed based on the boundary detection result; the viewpoint selecting module is used for selecting an optimal target viewpoint from target viewpoints to be executed by utilizing the door and window semantic information; and the path planning module is used for carrying out exploration path planning based on the optimal target viewpoint to obtain a corresponding exploration track, so that the unmanned platform explores the building space according to the exploration track.

Description

Autonomous exploration method and system for building space Technical Field The disclosure relates to the technical field of unmanned platform autonomous exploration, in particular to a building space autonomous exploration method and system. Background In recent years, with rapid development of robots and artificial intelligence technology, unmanned platforms such as service robots, sweeping robots, intelligent unmanned aerial vehicles, etc. have been increasingly used in urban building spaces to replace humans to perform specific works. The unmanned platform reduces the labor burden and the cost on one hand and improves the application efficiency and the convenience on the other hand. However, the unmanned platform applied to the urban building space has the problems that more human intervention is needed in the operation process, the autonomous level and the intelligent degree are limited, and the completeness and the comprehensiveness are limited due to the fact that the problem is modeled in the two-dimensional space to be solved by greatly depending on the structural characteristics of the building space. These problems make it difficult for current unmanned platforms to perform complex tasks successfully within a building space. To solve these problems, autonomous exploration of unknown space is becoming a hotspot problem in the robot field. The autonomous exploration of the unknown space means that no human intervention and prior information are needed, the unmanned platform explores and perceives the specific space by means of the sensor, the exploration route is planned autonomously, and meanwhile, the three-dimensional structure of the space is reconstructed and the target object is identified. In the task of autonomous exploration of an unknown space, how to select an optimal target point in an unknown environment is the core of problem solving. The boundary-based unknown space exploration method is a mainstream method in a current unknown space autonomous exploration algorithm, wherein the boundary is defined as a boundary between a known free space and an unknown space, and in the boundary-based exploration, the unmanned aerial vehicle is driven to move to an unexplored boundary area by detecting boundary points in real time and taking the boundary nearest to the current position of the unmanned aerial vehicle as a target point, so that the coverage of the unknown space is realized. In addition, unlike the above-described methods that select the boundary between the known free space and the unknown space as the boundary, some improved methods select the boundary within the sensor field of view that minimizes the speed change of the drone to maintain a higher exploration speed. Still other improved methods integrate the distance of the drone to the boundary, the size of the boundary region, and the steering angle (i.e., the cost of rotation) by introducing an improved cost function, thereby more rationally selecting the target boundary and reducing unnecessary steering and return visits. The unknown space exploration method based on sampling is another mainstream unknown space autonomous exploration method, and is characterized in that candidate viewpoints are randomly sampled in an exploration space, and the viewpoint with the highest information gain is selected as the next target point. The method first builds a fast random spanning tree (Rapidly-exploring Random Tree, RRT) in a known free space, generates a series of search points, then, for each search point generated, calculates its coverage of an unknown region (the number of voxels that are unknown and not occluded within the sensor field of view) and a weighted value of the collision-free path distance from the current location to the search point, the weighted value being defined as the information gain, the point with the highest information gain being selected as the target, and derives the traversable path from the RRT. This approach was originally introduced in the Next-Best-view planner (Next-Best-VIEW PLANNER, NBVP) and was further improved by gaussian belief propagation (Gaussian Belief Propagation, GBP) and motion primitive-based planners (Motion Primitive-Based Planner, MBP). The GBP constructs a global topological map in the exploration process, and if the current area is fully explored or the unmanned aerial vehicle enters the dead peucedanum, the method can find a path on the global map and redirect the unmanned aerial vehicle to the unexplored area. MBP builds RRT using motion primitives (Motion primitives) and generates smooth trajectories for unmanned aerial vehicles to execute. The prior art also provides an intuitive method, namely a method for fusing an unknown space exploration method based on a boundary and an unknown space exploration method based on sampling, wherein candidate positions are sampled in a boundary area, the value of each candidate position is estimated through a utility function, and finally, the o