Mobile robot laser radar mapping, light-transmitting environment false object identification and map correction method

Abstract

A method for identifying false objects and correcting maps in a transparent environment generated by a LiDAR mapping system for mobile robots involves dilating obstacle points in the original occupied grid map generated by the LiDAR mapping system to connect discontinuous obstacle points, achieving initial separation between the transparent map and the real map. Then, a region growing method is used to grow from the origin of the original occupied grid map outwards, distinguishing between the unknown map, the transparent map, and the real map. Through the synergistic combination of the above morphological preprocessing and region growing segmentation strategies, this invention can effectively distinguish and eliminate those transparent false regions that are spatially isolated from the main body of the real map, ultimately outputting an environment map with accurate geometric structure and reliable semantic information, providing high-quality environmental representation support for subsequent higher-level tasks such as robot navigation and obstacle avoidance.

Description

Technical Field

[0001] This invention belongs to the field of intelligent robot technology, and specifically relates to a method for identifying false objects and correcting maps in a transparent environment using LiDAR mapping for mobile robots. Background Technology

[0002] LiDAR, with its centimeter-level ranging accuracy and robustness to illumination, has become the mainstream sensing unit in the field of simultaneous localization and mapping (SLAM) for mobile robots. The grid map generated based on LiDAR modeling provides a crucial environmental geometric representation foundation for higher-level control tasks such as robot path planning, autonomous navigation, and dynamic obstacle avoidance. The map's accuracy directly determines the reliability and operational safety of the robot.

[0003] In complex indoor and outdoor scenarios involving architectural structures with optical penetration characteristics, such as glass curtain walls and translucent windows, LiDAR mapping faces technical challenges. Because laser beams can penetrate glass surfaces, some laser energy illuminates background objects behind the glass (such as exterior building facades or furniture in adjacent indoor spaces), generating reflected echoes. This results in the radar receiving point cloud data containing false obstacle information from outside the current space. This light transmission effect introduces numerous false target features into the constructed map, severely distorting the geometric representation of the real environment. This leads to risks such as robot localization drift, path planning failure, and even collisions, significantly limiting the effectiveness of LiDAR SLAM systems in complex architectural scenarios.

[0004] To address the aforementioned technical bottlenecks, existing research primarily focuses on the accurate detection and semantic annotation of glass surfaces, resulting in two typical technical approaches: One type is the single-mode detection method based on laser echo characteristics, represented by the scheme proposed by Wang et al. Its core mechanism lies in the fact that the reflection intensity reaches its peak when the laser beam is perpendicularly incident on the glass surface, and then monotonically decreases with increasing incident angle. Therefore, when the mapping robot moves along the glass facade, the laser echo intensity sequence received by the radar exhibits a characteristic fluctuation pattern of first increasing and then decreasing. By designing an intensity sequence analysis algorithm with an adaptive threshold, real-time identification of the glass surface can be achieved, and it can be marked as an obstacle region with specific semantics during map construction, thereby eliminating interference from false point clouds.

[0005] Another type is the multimodal perception method that integrates visual and depth information, represented by the technical solution proposed by Zhao et al. This method simultaneously acquires color and depth images of the environment using an RGB-D camera. First, it performs visual SLAM modeling based on the ORB-SLAM algorithm to obtain a sparse map point cloud of the scene and the camera's motion trajectory. Then, it uses the PSPNet deep learning network to perform pixel-level semantic segmentation on the RGB images, achieving two-dimensional semantic recognition of the glass region. Finally, by fusing large-scale planar geometric features extracted from the depth images, it performs cross-modal registration of the two-dimensional semantic segmentation results with the three-dimensional point cloud structure to generate an environmental map containing accurate semantic information of the glass surface, providing a reliable environmental representation for subsequent robot decision-making.

[0006] In summary, the main approach of existing technologies is the identification and semantic annotation of obstacles such as glass. Whether based on the physical properties of laser reflection from the glass surface using lidar or on the extraction of visual appearance features of the glass, such as transparency and texture differences, using visual sensors, the aim is to achieve accurate identification and localization of glass in complex environments based on the response patterns of glass to specific sensor signals. This has led to two representative technical solutions: single-modal detection methods based on lidar echo intensity characteristics and multi-modal perception methods based on visual semantic segmentation. Although these solutions have achieved effective identification of glass areas in specific scenarios, they still have several inherent limitations.

[0007] For detection methods based on laser echo intensity information, the core of this method relies on the physical characteristic that the reflection intensity reaches its peak when the laser beam is nearly perpendicularly incident. If the laser scanning angle deviates from the vertical direction and exceeds the threshold range, it is easy to miss detections in glass areas. Furthermore, real-time monitoring and analysis of the fluctuation characteristics of the laser echo intensity sequence not only increases the computational load of the synchronous localization and map building process, potentially affecting the system's real-time response performance, but also introduces the risk of false detections in glass areas due to interference from other highly reflective materials in the environment (such as mirrors, polished metal surfaces, etc.), reducing the accuracy of semantic annotation. Multimodal methods based on visual semantic segmentation depend on the acquisition quality of RGB images and the visual texture features of the glass surface. Under complex lighting conditions such as uneven illumination, overexposure, underexposure, or strong specular reflection on the glass surface, the effective information content of the image decreases significantly, weakening the feature extraction ability of the deep learning model and consequently significantly reducing the accuracy of glass area recognition. Furthermore, in scenarios where the glass surface lacks texture (such as plain transparent glass) or where external objects are difficult to distinguish from the glass background, the semantic segmentation boundaries of this method are prone to blurring or shifting, leading to misjudgments or omissions of glass areas, thus limiting its applicability. Existing solutions share a common limitation: they all focus on the detection and semantic annotation of glass, without addressing the core issue caused by the light-transmitting characteristics of glass—map contamination. Map contamination refers to the phenomenon where a laser beam, after penetrating the glass medium, illuminates background objects outside the current workspace and generates reflected echoes, thus introducing false obstacle information that does not actually exist into the point cloud data. Current mainstream methods, after completing the annotation of glass areas, typically lack automated identification and adaptive filtering mechanisms for this type of false point cloud information, and have not established a systematic method for correcting contaminated maps. This directly results in the final mapping results still retaining a large amount of geometrically distorted information, requiring manual post-processing. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, the present invention aims to provide a method for identifying false objects and correcting maps in transparent environments when using LiDAR mapping for mobile robots. This method addresses the problem of false obstacle point clouds and map distortion caused by the "transmission effect" in LiDAR mapping of transparent environments, including glass. Based on the principle of spatial connectivity, it achieves automated map correction. Through 3D point cloud data fusion and neighborhood search strategies, it achieves accurate segmentation between the real map and the map containing transmission errors, providing accurate maps for navigation and control of robots in complex scenarios.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying false objects and correcting maps in a light-transmitting environment using LiDAR mapping for mobile robots includes the following steps: Step 1: Dilate the obstacle points in the original occupied grid map generated by the laser mapping system, connect the discontinuous obstacle points, and achieve the initial separation between the transparent map and the real map. Step 2: Using the region growing method, the map is grown from the original mapping origin of the original occupied grid map outwards to distinguish between the unknown map, the transparent map, and the real map.

[0010] The principle of this invention is that the real physical environment has the characteristics of spatial continuity and connectivity, while the false map area formed by laser penetrating glass has a spatial distribution that is isolated from the main body of the real map. Based on this essential difference, accurate segmentation of the real area and the false area can be achieved.

[0011] Therefore, the core objective of step 1 of this invention is to enhance the topological structural features of the map. Although the glass medium causes light transmission, entities such as the window frame and edges of the glass can be detected by LiDAR and generate realistic obstacle point clouds. Only such point clouds exhibit a sparse and discontinuous distribution in the original map. This invention uses circular structural elements to perform an expansion operation on the occupied grid. Through morphological transformation, adjacent occupied grid points belonging to the same real physical structure can be connected and merged, strengthening the spatial integrity of the real environment area.

[0012] Subsequently, the core design basis of step 2 of this invention is the reliable prior knowledge that "the mapping origin (the robot's initial mapping position) must belong to the real map area." Using this origin as the initial seed point for region growth, and through preset growth rules that conform to physical space constraints, all real environment areas connected to the seed point are automatically marked. The specific growth rules are defined as follows: if the current target grid is an empty area, growth expansion to adjacent empty and occupied grids is allowed; if the current target grid is an occupied area, growth expansion is only allowed to adjacent occupied grids. This growth process essentially simulates the real environment range that the robot can physically reach from its initial position, ensuring the consistency between the growth area and the real physical space.

[0013] By combining the above-mentioned morphological preprocessing with the region growing segmentation strategy, this invention can effectively distinguish and eliminate those translucent false regions that are spatially isolated from the real map subject, and finally output an environmental map with accurate geometric structure and reliable semantic information.

[0014] Compared with existing technologies, this invention can automatically clean up erroneous maps caused by light transmission through glass surfaces in laser mapping without manual intervention, significantly improving map processing efficiency. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the process of this invention.

[0016] Figure 2 This is a schematic diagram of the original occupying point and the circular expansion core during the occupation point expansion process.

[0017] Figure 3 This is a schematic diagram of the expanded grid after the expansion process of the occupied point.

[0018] Figure 4 This is an embodiment of the present invention. A schematic diagram of the marked grid in the initial state of a simplified map.

[0019] Figure 5 yes Figure 4 A schematic diagram of the marked grid representing the growth status of blank points on a simplified map.

[0020] Figure 6 yes Figure 4 A schematic diagram of the marked grid representing the growth status of occupied points on a simplified map.

[0021] Figure 7 yes Figure 4 A schematic diagram of the final marked grid of the simplified map.

[0022] Figure 8 yes Figure 4 The original map corresponding to the simplified map.

[0023] Figure 9 yes Figure 4 The simplified map is the final map after being processed by the light transmission method of this invention.

[0024] Figure 10 This is the original indoor map in an embodiment of the present invention.

[0025] Figure 11 Yes Figure 10 An image showing the effect of removing light-transmitting parts from an indoor map.

[0026] Figure 12 This is the original map of the outdoor corridor in this embodiment of the invention.

[0027] Figure 13 Yes Figure 12 An image showing the effect of removing the light-transmitting parts from the outdoor corridor map. Detailed Implementation

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples.

[0029] During laser mapping, the glass surface causes light transmission. However, the rough surfaces such as the window frames embedded between the glass panes, the slight diffuse reflection of the glass itself, and the specular reflection when the light is incident perpendicularly, allow the laser mapping to still detect some obstacle points. These obstacle points are usually discontinuous. This invention first dilates the obstacle points in the map to connect the discontinuous obstacle points, achieving an initial separation between the translucent map and the real map. After the initial separation, a region growing method is used to automatically distinguish between the translucent and real parts.

[0030] refer to Figure 1 As shown, the specific process of the present invention is as follows: Step 1: Input the original occupied grid map.

[0031] Input the original occupied raster map generated by the laser mapping system. The size is (Width × Height). Originally occupying each grid cell in the grid map. There are three state values: : Indicates that the grid is occupied by an obstacle (such as a detected window frame or wall), and is defined as an occupied point.

[0032] : Indicates that the grid is an empty area (without obstacles), defined as a blank point.

[0033] : Indicates that the state of the grid is unknown (area not scanned by the laser, such as the back of a wall), and is defined as an unknown point.

[0034] Step 2: Perform morphological dilation on the portion of the map that it occupies.

[0035] To connect the discrete occupied points caused by the glass light transmission effect, the original map was modified. The expansion operation is performed on the occupied area (the grid with state 1) within the structure. A circular structuring element is defined as the expansion kernel. ,like Figure 2 As shown, its mathematical expression is: in, This represents the offset of the internal grid cells relative to the central grid cell. The expansion radius is calibrated based on the spacing of the points occupied in the actual environment. Represents integers. Obstacle points are expanded to obtain a new map. ,like Figure 3 As shown, the assignment rules for each grid cell are as follows: After the expansion process, the obstacles that were originally discontinuous due to the light-transmitting objects are connected, achieving the initial separation of the real map and the light-transmitting map.

[0036] This step ensures that neighboring isolated occupancy points are connected into a continuous area, laying the foundation for subsequent area growth.

[0037] In the definition of this invention, the origin of the original occupied grid map belongs to the real map, which is a core prior condition. Each pixel in the occupied grid map is divided into three states: blank, occupied, and unknown. By growing outward from the origin, the unknown part, the transparent part, and the real part can be distinguished.

[0038] Step 3: Execute the region growing algorithm based on the origin of the map.

[0039] This step aims to start from a known real map area, i.e., the origin of the map. Starting from the origin, all reachable real grid cells are automatically marked, thus separating regions with light-transmitting fallacies. The region growing method grows outwards from the mapping origin to distinguish between unknown maps, light-transmitting maps, and real maps. Its implementation is as follows: Create a queue and add the origin of the graph. Perform a four-neighbor search on the head of the queue: if the head of the queue is an empty point, add the empty points and occupied points in its four-neighbors to the queue; if the head of the queue is an occupied point, only add the occupied points in its four-neighbors to the queue. Remove the head of the queue from the queue and repeat the above operation until the queue is empty. Points that have entered the queue in the original grid map are marked as real map points, while unmarked points are unknown map points or transparent map points.

[0040] The specific process of creating the queue and adding the origin of the graph can be described as follows: Step 31: Initialization, define the tag matrix and pending queue ,in Elements in: The origin of the map Add to the queue and mark as a real map: .

[0041] exist Figure 4 In the simplified map shown, white represents blank grid cells, black represents occupied grid cells, and gray represents unknown grid cells; the origin of the real map is marked therein. Dyed orange; located white grid These are the points where light is transmitted. Marker matrix. and search queue They are respectively: The specific process of performing a four-neighbor search on the head of the queue can be described as follows: Step S32: Loop processing.

[0042] When the queue is not empty, retrieve the first grid cell. And check its four neighbors: superimpose the offset set on the original coordinates: ; If the first point of the queue is blank, that is... Add unmarked blank or occupied points in its four neighboring areas to the queue and mark them as real map points.

[0043] like Figure 5 Squadron First Point For blank points, grow to their four neighborhoods. After updating the label matrix and search queue, they are as follows: If the first player in the team occupies a point, that is... Add the unmarked occupied points in its four neighboring areas to the queue and mark them as real map points.

[0044] like Figure 6 Squadron First Point As for the occupied points, the algorithm only grows upwards and downwards towards the occupied points. After updating, the marker matrix and search queue are as follows: The specific judgment conditions are as follows: for like like The condition for terminating the loop is: When queue When the region is empty, the region growth ends, and the label matrix is... Completely marked real map With unknown / transparent maps The labeling results of the simplified grid are as follows: Figure 7 As shown, the final tag matrix is: Step 4: Clean up the graticule that was not marked as real.

[0045] After the loop ends, the unmarked graticule cells are cleaned up, resulting in the final map. Defined as: In a raster map, points that have entered the queue are marked as real map points, while unmarked points are either unknown or transparent parts. This indicates an unmarked grid, which can be directly removed according to the robot's map requirements.

[0046] like Figure 8 Grid in the original map For the blank spots created by light transmission, in Figure 7 Columns one through three are marked as real maps, retaining their original occupied or blank states; the fourth column is unmarked, and its grid cells are assigned an unknown state. Ultimately... Figure 9 Misconceptions caused by light transmission It was eliminated.

[0047] Step 5: Output the map.

[0048] In summary, this invention achieves accurate and automated segmentation of real and translucent maps in laser mapping by combining image morphology processing and region growing strategies, thus eliminating reliance on manual processing.

[0049] The specific experimental verification of this invention involves mapping an indoor environment containing glass, resulting in a map as shown below. Figure 10 As shown, there are radial light-transmitting grids on the left side of the map. After processing, in Figure 11 Rays produced by light transmission were successfully eliminated. When drawing in a long outdoor corridor, the light transmission through the glass windows of the rooms on both sides of the corridor created numerous fan-shaped blank error grids, such as... Figure 12 As shown. After processing, in Figure 13 The fan-shaped error grid has been largely eliminated, leaving only the correct main corridor map.

Claims

1. A method for identifying false objects and correcting maps in a light-transmitting environment using LiDAR mapping for mobile robots, characterized in that... Includes the following steps: Step 1: Dilate the obstacle points in the original occupied grid map generated by the laser mapping system, connect the discontinuous obstacle points, and achieve the initial separation between the transparent map and the real map. Step 2: Using the region growing method, the map is grown from the original mapping origin of the original occupied grid map outwards to distinguish between the unknown map, the transparent map, and the real map.

2. The method for identifying false objects and correcting maps in a transparent environment using LiDAR mapping for mobile robots according to claim 1, characterized in that, Each grid in the original occupied grid map There are three state values: : Indicates that the grid cell is occupied by an obstacle, defined as an occupied point; This indicates that the grid is an empty area with no obstacles, and is defined as a blank point; : Indicates that the state of the grid is unknown, that is, the area not scanned by the laser, which is defined as an unknown point.

3. The method for identifying false objects and correcting maps in a transparent environment using lidar mapping for mobile robots according to claim 2, characterized in that, A new map is obtained by dilating obstacle points in the original occupied grid map generated by the laser mapping system. The assignment rules for each grid cell are as follows: in, For the expansion kernel, a circular structural element is used, mathematically expressed as: in, This represents the offset of the internal grid cells relative to the central grid cell. The expansion radius is calibrated based on the spacing of the points occupied in the actual environment. Representing integers, after dilation, the obstacles that were originally discontinuous due to the light-transmitting objects are connected, achieving the initial separation of the real map and the light-transmitting map.

4. The method for identifying false objects and correcting maps in a transparent environment using LiDAR mapping for mobile robots according to claim 1, characterized in that, Step 2, the region growing method, is as follows: Create a queue and add the origin of the graph. Perform a four-neighbor search on the head of the queue: if the head of the queue is an empty point, add the empty points and occupied points in its four-neighbors to the queue; if the head of the queue is an occupied point, only add the occupied points in its four-neighbors to the queue. Remove the head of the queue from the queue and repeat the above operation until the queue is empty. Points that have entered the queue in the original grid map are marked as real map points, while unmarked points are unknown map points or transparent map points.

5. The method for identifying false objects and correcting maps in a transparent environment using lidar mapping for mobile robots according to claim 4, characterized in that, The creation of the queue and the addition of the graph origin are represented as follows: Define the tag matrix and pending queue ,in Elements in: The origin of the map Add to the queue and mark as a real map: 。 6. The method for identifying false objects and correcting maps in a transparent environment using LiDAR mapping for mobile robots according to claim 5, characterized in that, The four-neighbor search performed on the head of the queue is represented as: When the queue is not empty, retrieve the first grid cell. And check its four neighbors: superimpose the offset set on the original coordinates: ; If the first point of the queue is blank, that is... Add unmarked blank or occupied points in its four neighboring areas to the queue and mark them as real map points; If the first player in the team occupies a point, that is... : Add the unmarked occupied points in its four neighboring areas to the queue and mark them as real map points; When queue When the region is empty, the region growth ends, and the label matrix is... Completely marked real map With unknown / transparent maps .

7. The method for identifying false objects and correcting maps in a transparent environment using lidar mapping for mobile robots according to claim 5, 6, or 7, is characterized in that... After the loop ends, the unmarked graticule cells are cleaned up, resulting in the final map. Defined as: This indicates an unmarked grid, which will be removed directly based on the robot's map requirements.

Images