A point cloud ground real-time segmentation method and system for post-disaster buried space

Abstract

The application relates to the technical field of laser radar point cloud processing, in particular to a point cloud ground real-time segmentation system and method for post-disaster buried space. The system mainly comprises three core modules, which are, in sequence, a real-time monitoring and session management module responsible for data stream access and task scheduling, a parameter centralized configuration module capable of flexibly adjusting algorithm performance to adapt to different environments, and an end-to-end processing module for executing core segmentation tasks. Through timestamp subfolder real-time monitoring, parameter unified management and Patchwork++ ground segmentation, the ground points (ground) and unground points (unground) are quickly separated. The system design fully considers the resource-limited application scenario and can be deployed and run on an embedded computing power board. The application can output a ground segmentation result with high real-time performance and high robustness in a complex underground environment, is suitable for emergency rescue scenes in post-disaster buried space, and has high engineering practical value.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence technology, specifically a method and system for real-time segmentation of point cloud ground for post-disaster buried space. Background Technology

[0002] With the large-scale development and utilization of urban underground space in my country, facilities such as subway tunnels and underground parking lots often face severe structural damage, irregular spatial forms, and continuously changing environments in post-disaster burial scenarios, posing significant challenges to emergency rescue. Under such complex conditions, rescuers urgently need technical means to quickly and accurately acquire information about the internal structure of the burial area to support rescue decision-making and operational safety. While lidar, as the primary sensor, can provide high-precision three-dimensional geometric information, the raw point cloud data is highly mixed with ground and non-ground points, easily generating significant redundant interference and unnecessary computational burden for subsequent mapping, positioning, and 3D reconstruction. This severely restricts the system's real-time performance and overall perception efficiency, making it difficult to meet the urgent needs of post-disaster rescue for refined spatiotemporal perception and sensor-computation collaboration.

[0003] A search revealed that the invention patent with publication number CN112802096A proposes a ground segmentation method for lidar point clouds. This method achieves ground point extraction through region growing and multi-feature fusion. Although it improves the segmentation accuracy to a certain extent in conventional structured scenarios, it is mainly aimed at single-frame point cloud processing and lacks a real-time monitoring and session management mechanism for continuous scanning data streams. It is not adaptable enough to problems such as trajectory drift, smooth wall reflection, and sparse geometric features in unstructured buried spaces.

[0004] Another invention patent with publication number CN113569922A proposes a ground segmentation method and device based on point cloud. It uses rasterized projection combined with height statistics and plane fitting to achieve ground separation. Although it has a certain computational efficiency, it still has obvious defects in terms of scattered parameter configuration, long-term stable operation of embedded computing platform and system integration.

[0005] The aforementioned existing technologies generally suffer from the following shortcomings in post-disaster buried space applications: First, they struggle to adapt to continuous multi-source data streams stored by timestamps on equipment platforms, lacking effective real-time monitoring and session management mechanisms, leading to data processing prone to out-of-order processing or interruptions; second, parameter configurations are relatively scattered, making it difficult to quickly adapt to different complex underground scenarios; and third, system integration is low, with poor coordination with multi-sensor data fusion processes, making it difficult to form a complete and efficient "map first, segmentation then reconstruction" processing link. A more efficient and stable end-to-end processing link is needed.

[0006] Therefore, there is an urgent need for a point cloud ground real-time segmentation system and method that can achieve real-time data monitoring, centralized parameter management and tight coupling with multiple sensors, so as to significantly improve the structured perception capability of buried space after disaster and the efficiency of rescue operations. Summary of the Invention

[0007] This invention aims to address the problems of low efficiency in real-time monitoring of point cloud ground segmentation in post-disaster buried spaces, scattered parameter configuration, poor adaptability to embedded deployment, and insufficient continuity of data stream processing in existing technologies. It provides a method and system for real-time ground segmentation of point clouds in post-disaster buried spaces. Through an innovative real-time monitoring and session management mechanism, a centralized parameter configuration module, and an efficient end-to-end processing pipeline, this system achieves efficient ground and non-ground separation of LiDAR point clouds, providing reliable technical support for fine-grained perception and subsequent analysis of complex underground environments.

[0008] The main technical solutions of this invention include the following core innovative designs: First, a real-time monitoring and session management mechanism is proposed, which can automatically identify point cloud files in subfolders named with timestamps and maintain an independent session state for each folder. This mechanism effectively solves the problem of orderly processing of continuous, multi-batch data streams generated by the equipment platform in handheld or backpack scanning modes through polling the directory, dynamic state dictionary management, and silent timeout processing strategies, ensuring the continuity and stability of data processing. Second, a centralized parameter configuration module is constructed, which uniformly manages and injects key parameters such as sensor installation height, area division parameters, iteration count, distance threshold, and noise removal switch, supporting rapid and flexible optimization according to different underground scene characteristics, significantly reducing the difficulty of on-site deployment and algorithm optimization. Third, a complete end-to-end processing pipeline is designed, covering point cloud reading and intensity calculation, execution of core ground segmentation algorithms, construction and persistent storage of ground and non-ground point clouds, and fine-grained performance statistics. This pipeline supports optional color information processing, which can be flexibly balanced according to actual computing power and accuracy requirements, further improving the engineering practicality of the system.

[0009] This invention uses Patchwork++ as the core segmentation algorithm, based on the Concentric Zone Model (CZM). It divides the input point cloud into multiple rings and sectors using polar coordinates, and performs Region-wise Ground Plane Fitting (R-GPF) within each local region. Combined with Adaptive Ground Likelihood Estimation (A-GLE), it achieves dynamic threshold adjustment, thus exhibiting strong robustness to complex, unstructured underground environments. Furthermore, this invention introduces a fine-grained performance statistics mechanism, which can record the time consumption and point cloud quantity at each processing stage in real time, providing crucial data support for resource monitoring, performance optimization, and subsequent algorithm iteration on embedded platforms.

[0010] This invention significantly improves the real-time performance, stability, and scene adaptability of point cloud ground segmentation in post-disaster buried space through the organic combination and collaborative work of the above modules. The single-frame processing time meets the real-time requirements of embedded platforms, providing high-quality structured ground and non-ground point cloud data for subsequent spatial analysis, rescue route planning, and decision support, and has broad engineering application prospects. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 The diagram shows the overall architecture of the system of this invention, illustrating the complete processing flow from single-frame point cloud data acquisition at the sensing end, multi-frame point cloud fusion and storage at the host end, to real-time segmentation of the point cloud and ground at the computing board. Figure 2 The flowchart for real-time monitoring and session management details the entire process, including timestamp subfolder identification, session state initialization, file processing, and silent timeout handling. Figure 3 This diagram illustrates the end-to-end point cloud processing pipeline, showcasing the main steps of point cloud reading, Patchwork++ segmentation, and the construction and storage of ground and non-ground point clouds. Figure 4 This is a schematic diagram of the core segmentation mechanism of Patchwork++, illustrating the process of concentric region model partitioning, regional ground plane fitting, and adaptive threshold determination. Figure 5 This is a segmentation effect diagram of a typical complex underground scene, which intuitively presents the separation results of white ground point cloud and colored non-ground point cloud. Detailed Implementation

[0012] This invention provides a real-time ground segmentation system and method for post-disaster buried spatial point clouds, the specific implementation of which is combined with Figures 1 to 5 Please provide a detailed explanation.

[0013] like Figure 1 As shown, the system of this invention is mainly divided into three layers: the sensing end, the host end, and the computing board end. The sensing end is responsible for collecting single-frame point cloud data, including the original point cloud of the LiDAR, intensity information, and RGB color data; the host end completes the fusion and storage of multiple frame point clouds to form a continuous data sequence named with timestamps; the computing board end serves as the core computing unit, deploying a real-time monitoring module, a centralized parameter configuration module, and a core processing module to achieve efficient real-time segmentation of the point cloud ground.

[0014] The operation flow of the real-time monitoring module is as follows: Figure 2 As shown. After system startup, it continuously polls the point cloud file directory. When a new timestamp subfolder is detected, the session state is initialized, and the corresponding segmentation result output directory is created. Subsequently, the .pcd files are sorted according to their modification time, and the core processing modules are called sequentially for segmentation. After a silent timeout, the system automatically writes to the runtime log and clears the session state to ensure stable long-term operation.

[0015] Core processing modules such as Figure 3 As shown. First, the .pcd file is read using Open3D and intensity calculations are performed. Then, Patchwork++ is called to perform ground segmentation, extracting ground and non-ground indices, and constructing and saving _ground.pcd and _unground.pcd files respectively. Simultaneously, the time consumption and point cloud quantity statistics for each stage are recorded. This invention uses a concentric zone model (CZM) to divide the input point cloud into multiple rings and sectors according to polar coordinates. Region-wise ground plane fitting (R-GPF) is performed within each local region. The core calculation formula is the point-to-plane distance residual: Where n is the local plane normal vector calculated by PCA, For the point cloud points to be classified, Use the regional reference point. Combine Adaptive Ground Likelihood Estimation (A-GLE) to dynamically adjust the elevation and flatness thresholds. The time was determined to be a ground point.

[0016] The parameter centralized configuration module injects all parameters related to segmentation performance into the command-line interface through the `add_segmentation_args` function, and constructs a Patchwork++ instance through the `build_patch_from_args` function. These parameters include `sensor_height` (sensor installation height, typically set to around 1.5m), `th_seeds` (initial seed point height threshold), `th_dist` (point-to-plane distance threshold), `num_zones` (number of concentric region partitions), `num_rings_of_interest` (number of interest rings), and `enable_RNR` (reflection noise removal switch). In practical applications, users can flexibly adjust these parameters according to specific scenarios.

[0017] like Figure 4 As shown, this invention uses the Patchwork++ core segmentation mechanism to finely partition the input point cloud through a concentric region model, and completes seed point extraction, plane fitting and adaptive threshold classification in each region.

[0018] like Figure 5 As shown, in typical underground scenario tests, this invention can clearly separate the original point cloud into ground point cloud (continuous flat area) and non-ground point cloud (walls, etc.), effectively reducing the amount of redundant data in subsequent processing, and providing high-quality structured data support for rescuers to quickly grasp the internal structure of the buried area, potential risk points and passable paths.

Claims

1. A real-time ground segmentation system for point clouds buried in post-disaster environments, characterized in that, It includes a real-time monitoring module, a centralized parameter configuration module, and a core processing module; the real-time monitoring module is used to poll the point cloud file directory and identify timestamp subfolders to maintain an independent session state; the centralized parameter configuration module is used to uniformly manage Patchwork++ segmentation-related parameters; the core processing module is used to implement point cloud reading, ground segmentation, and output of ground.pcd and unground.pcd files.

2. The system according to claim 1, characterized in that, The real-time monitoring module scans the directory through a polling mechanism, sorts the .pcd files by file modification time, and uses a session state dictionary to manage the output directory, generated file list, and processing records of each timestamp subfolder. It also supports automatic writing to the running log and clearing the session state after a silent timeout.

3. The system according to claim 1, characterized in that, The core processing module includes the following steps: reading the .pcd file through Open3D and calculating the intensity value; calling Patchwork++ to perform ground segmentation and extracting ground and non-ground indices; constructing and saving the _ground.pcd and _unground.pcd files respectively, while recording the time consumption statistics of each processing stage.

4. The system according to claim 1, characterized in that, The parameter centralized configuration module uses the add_segmentation_args and build_patch_from_args functions to uniformly inject and configure key parameters such as sensor_height, th_seeds, th_dist, and num_zones, supporting rapid adaptation to different underground scenarios.

5. The system according to claim 1, characterized in that, The core processing module uses a concentric region model (CZM) to partition the point cloud, and calculates the point-to-plane distance residual within each local region through regional ground plane fitting (R-GPF). ;in, The local plane normal vector is obtained by PCA. For the point cloud points to be classified, Used as a regional reference point.

6. The system according to claim 1, characterized in that, The system is deployed on an embedded computing board and supports real-time processing of timestamp point cloud data output from the host, achieving millisecond-level single-frame ground segmentation.

7. A method for real-time ground segmentation of point clouds buried in post-disaster environments, characterized in that... Includes the following steps: S1: Monitor the point cloud file directory in real time and identify timestamp subfolders, and initialize the session state; S2: Load the centrally configured parameters to construct a Patchwork++ instance; S3: Read the .pcd file and perform ground segmentation; S4: Build and save ground point cloud and non-ground point cloud files; S5: Record processing statistics and write them to the runtime log after a silent timeout.

8. The method according to claim 7, characterized in that, The ground segmentation step uses a concentric region model to partition the area, and calculates the residual distance from the point to the plane by fitting the regional ground plane within each region. The ground point classification is then completed by combining the residual distance with an adaptive threshold.

9. The method according to claim 7, characterized in that, The method supports optional processing of color information and records the time consumption and point cloud quantity statistics at each stage during the processing.

10. The system or method according to claim 1 or 7, characterized in that, It is suitable for real-time point cloud ground segmentation tasks in complex scenarios such as subway tunnels and underground parking lots after disasters.

Images