A laser radar measurement method and device based on multiple platforms

Abstract

The embodiment of the present invention provides a multi-platform laser radar measurement method and device, which realizes the application of the same device in multiple modes such as drones, vehicles, backpacks, etc. by modifying and adapting the laser radar equipment. The system integrates multiple sensors such as laser radar, Beidou navigation, inertial measurement unit, etc., and combines data processing technology. It is suitable for the complex scene of karst landforms and solves the technical problem that traditional single equipment is difficult to cope with high-precision measurement in complex environments. Through this technology, the efficiency of surveying and mapping operations and the accuracy of data acquisition can be improved to meet the measurement needs in multiple scenarios.

Description

Technical Field

[0001] The present invention relates to the field of laser radar measurement technology, and in particular to a laser radar measurement method and device based on multiple platforms, which is suitable for topographic mapping in karst landforms. Background Art

[0002] LiDAR measurement technology is widely used in the field of topographic surveying and mapping due to its high precision and high efficiency. At present, common LiDAR measurement platforms include UAV platforms, vehicle-mounted platforms and walking measurement platforms. However, in complex terrain conditions such as karst landforms, the existing technology has the following problems:

[0003] The measurement method of a single platform is difficult to adapt to the complex terrain characteristics of karst landforms, resulting in insufficient integrity and accuracy of data collection; in addition, the existing point cloud data processing methods have limited ability to identify and process noise points, and are prone to misjudgment, especially under complex terrain conditions. Summary of the invention

[0004] In view of this, the present application provides a multi-platform based laser radar measurement method and device, which solves the problems in the prior art of poor adaptability of single platform measurement methods, low point cloud data processing accuracy, and insufficient multi-platform data collaborative processing capabilities.

[0005] The embodiment of the present application provides a multi-platform based laser radar measurement method, which is applied to the measurement of a laser radar measurement host in karst landforms. The laser radar measurement host includes a laser radar sensor, a Beidou navigation system and an inertial measurement unit. The method includes:

[0006] Collecting point cloud data obtained by the laser radar measurement host under the UAV platform, vehicle-mounted platform and walking measurement platform respectively, wherein the laser radar measurement host is installed on the rotor or fixed-wing bracket of the UAV platform, on the adjustable bracket of the vehicle-mounted platform and in a portable backpack device for walking measurement in a time-sharing manner;

[0007] Based on the real-time position data provided by the Beidou navigation system and the attitude data provided by the inertial measurement unit, the point cloud data is calibrated in space coordinates;

[0008] Inputting the calibrated point cloud data into a planning perceptual diffusion network to perform first noise removal, and using a secondary quadratic cutting plane method to perform classification processing on the point cloud data to divide the point cloud data into ground point cloud and non-ground point cloud;

[0009] A hierarchical structure processing module is used to perform hierarchical optimization of the ground point cloud and the non-ground point cloud to generate three-dimensional terrain data.

[0010] The calibrated point cloud data is input into a planning perceptual diffusion network for first noise removal, including:

[0011] Constructing terrain data segments according to the point cloud data, and inputting the terrain data segments into the planning perception diffusion network;

[0012] Predicting a first noise point in a local area of ​​the terrain data segment by using the planning-aware diffusion network;

[0013] Comparing the predicted first noise point information with the landform planning to verify the consistency between the first noise point information and the actual terrain features;

[0014] After the verification is completed, the first noise point in the calibrated point cloud data is removed.

[0015] Before removing the first noise point in the calibrated point cloud data, the method further includes:

[0016] Building a value function library based on historical surveying and mapping data and landform models, wherein the value function library stores point cloud noise processing strategies under different terrain scenarios;

[0017] According to the characteristics of the current point cloud data, the optimal point cloud noise processing strategy is retrieved and called from the value function library;

[0018] The optimal point cloud noise processing strategy is used to process the first noise point in the current point cloud data.

[0019] Before performing spatial coordinate calibration on the point cloud data, the method further includes:

[0020] Organizing the point cloud data collected at different time series and different spatial positions according to a time-space sequence;

[0021] Analyze point cloud data of different time series to extract long-term change trends of landforms;

[0022] Based on the point cloud data, the landform structure features are analyzed from different spatial perspectives to explore the three-dimensional morphological change rules of the landform structure features;

[0023] Based on the long-term change trend and the three-dimensional morphological change law, the point cloud data is preprocessed.

[0024] The point cloud data is classified and processed by using a secondary cutting plane method to divide the point cloud data into ground point cloud and non-ground point cloud, including:

[0025] Constructing multiple cutting planes based on the height information of the point cloud data and the spatial relationship between adjacent points;

[0026] Mark the point cloud that is closely connected to the ground and has a height below a first threshold as a ground point cloud;

[0027] Marking the point cloud separated from the ground or with a height higher than the first threshold as a non-ground point cloud;

[0028] The ground point cloud and the non-ground point cloud are stored separately.

[0029] The method further comprises:

[0030] Constructing a connection relationship graph of the non-ground point cloud;

[0031] Decomposing the nodes of the connection relationship graph using a deterministic network decomposition algorithm;

[0032] identifying different types of features in the non-ground point cloud based on height distribution, density, and spatial relationships;

[0033] analyzing tree and building features in the non-ground point cloud using a preset rule set;

[0034] The second noise point in the non-ground point cloud is processed using a maximum independent set algorithm.

[0035] The second noise point in the non-ground point cloud is processed using a maximum independent set algorithm, including:

[0036] Calculate the spatial independence parameters of point clouds and identify point clouds with discontinuous distribution;

[0037] Analyze the reflection intensity and isolation degree of the point cloud to determine the discrimination threshold of the second noise point;

[0038] Identifying and removing the second noise point based on the discrimination threshold;

[0039] Performing integrity assessment on the non-ground point cloud data after removing the second noise point;

[0040] The discrimination threshold of the second noise point is dynamically adjusted according to the integrity evaluation result.

[0041] A hierarchical structure processing module is used to perform hierarchical optimization of the ground point cloud and the non-ground point cloud, including:

[0042] Constructing a hierarchical structure according to the spatial distribution characteristics of the ground point cloud and the non-ground point cloud;

[0043] Divide the obstacle point cloud data into the first level;

[0044] The third noise data introduced by weather factors is classified into the second level;

[0045] Assign weights to nodes at each level based on the reflection intensity and density distribution of the point cloud;

[0046] Point cloud data is optimized based on node weights.

[0047] Optimize point cloud data based on node weights, including:

[0048] Connect and mark the continuously distributed obstacle point clouds to construct terrain contours;

[0049] Eliminate abnormal point cloud data based on node weight;

[0050] Dynamically adjust node weight coefficients according to different terrain features;

[0051] Generate a 3D model of the Earth's surface features.

[0052] The present application also provides a computer device, the computer device comprising:

[0053] at least one processor; and,

[0054] a memory communicatively connected to the at least one processor; wherein,

[0055] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the above-mentioned multi-platform-based lidar measurement method.

[0056] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned multi-platform-based lidar measurement method.

[0057] An embodiment of the present application also provides a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the above-mentioned multi-platform based lidar measurement method are implemented.

[0058] This application has the following technical effects:

[0059] 1. By adopting a multi-platform measurement method, the integrity and accuracy of data collection are improved, which can better adapt to the complex terrain characteristics of karst landforms;

[0060] 2. The noise processing method is combined with the planning perception diffusion network and the value function library to improve the quality of point cloud data;

[0061] 3. The secondary cutting plane method is introduced to classify point cloud data, which improves the classification accuracy of ground point clouds and non-ground point clouds;

[0062] 4. The hierarchical structure processing module is used to realize the collaborative processing of multi-platform data, thus improving the data processing efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments of the present disclosure or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 A schematic diagram of a multi-platform laser radar measurement method provided in an embodiment of the present application;

[0065] Figure 2 A schematic diagram of the structure of a planning-aware diffusion network provided in an embodiment of the present application;

[0066] Figure 3 A schematic diagram of the process of point cloud data classification processing provided in an embodiment of the present application;

[0067] Figure 4 A schematic diagram of the structure of a hierarchical structure processing module provided in an embodiment of the present application;

[0068] Figure 5 A schematic diagram of the structure of a multi-platform laser radar measurement device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is described below with reference to the accompanying drawings.

[0070] like Figure 1 In the embodiment of the present application, the process of the multi-platform laser radar measurement method includes:

[0071] S1: Collect the point cloud data obtained by the laser radar measurement host under the UAV platform, vehicle platform and walking measurement platform respectively; wherein the laser radar measurement host is installed on the rotor or fixed-wing bracket of the UAV platform, the adjustable bracket of the vehicle platform and the portable backpack device of the walking measurement in a time-sharing manner;

[0072] Among them, the laser radar measurement host is equipped with a high-precision laser radar sensor and can adapt to various platforms such as drones, vehicles and backpacks.

[0073] Inertial Navigation System (IMU): used to obtain device attitude information in real time and help improve the accuracy of measurement data.

[0074] GNSS system: Obtain the real-time location of the device through the Beidou or GPS positioning system to improve the geographic positioning accuracy of the data.

[0075] Specifically, the first step is to select high-precision lidar sensors and match them with GNSS and IMU equipment to ensure the stability of data acquisition under different operating platforms.

[0076] The second step is to design a drone adaptation module, including both rotary-wing and fixed-wing modes, and ensure that the equipment can work stably in the air, especially for large-area operations.

[0077] The third step is to design a vehicle-mounted bracket that can be adjusted in height to accommodate different vehicles and meet the needs of detailed ground mapping of urban streets, traffic, etc.

[0078] The fourth step is to design a portable backpack module to optimize the weight and center of gravity distribution of the equipment, ensure the comfort and measurement accuracy of the equipment when carried by a person, and make it suitable for walking operations in complex terrain.

[0079] Specifically, S1 involves the point cloud data collection process of the laser radar measurement host under multiple platforms. In the karst landform environment, due to the complexity and diversity of the terrain, it is difficult to obtain complete terrain data using a single platform measurement method. Therefore, this application adopts a multi-platform collaborative measurement method to collect point cloud data in different scenarios through drone platforms, vehicle-mounted platforms, and walking measurement platforms.

[0080] Specifically, the laser radar measurement host includes a laser radar sensor, a Beidou navigation system and an inertial measurement unit. Among them, the laser radar sensor can use a high-precision multi-line laser radar, such as a 64-line or 128-line laser radar sensor, with a scanning frequency of more than 100Hz, a scanning angle range of 360°, a measurement accuracy better than ±2cm, and a measurement distance of up to 200m. The laser radar sensor obtains the spatial position information of the target object by emitting a laser beam and receiving a reflected signal.

[0081] In the multi-platform measurement process, the LiDAR measurement host is installed in a time-sharing manner and installed on different measurement platforms according to the needs of the measurement scenario:

[0082] A1. UAV platform measurement: The laser radar measurement host is installed on the rotor or fixed-wing bracket of the drone through a special shock-absorbing device. Rotary-wing drones are suitable for fine-grained measurement of local areas and can hover in the air for fixed-point scanning; fixed-wing drones are suitable for rapid measurement of large areas. When measuring with a drone platform, the flight altitude is usually controlled between 50-150 meters, and the flight speed is adjusted according to the required point cloud density, usually between 5-20m / s.

[0083] A2. Vehicle-mounted platform measurement: The LiDAR measurement host is installed on an adjustable bracket on the vehicle-mounted platform, which has anti-vibration and automatic leveling functions. The height range of the adjustable bracket is 1.5-3 meters, which can be adjusted according to measurement requirements. The vehicle-mounted platform is mainly used to measure roads and surrounding areas. The driving speed is usually controlled between 10-40km / h to ensure that sufficient density of point cloud data is obtained.

[0084] A3. Walking measurement platform: The LiDAR measurement host is installed in a specially designed portable backpack device. The backpack device adopts ergonomic design, balanced weight distribution, shock absorption and waterproof functions. Walking measurement is mainly used for complex terrain areas that cannot be reached by vehicles, such as karst caves, steep mountains and other scenes. The walking speed is usually controlled between 2-4km / h.

[0085] During the data collection process, the laser radar measurement host adopts a unified data collection parameter setting:

[0086] 1. Scanning frequency: dynamically adjusted according to the movement speed of the measurement platform to ensure that the density of the point cloud data meets the measurement requirements;

[0087] 2. Data storage format: Use standard LAS format or PCD format to facilitate subsequent processing;

[0088] 3. Time synchronization: The time synchronization of each sensor is achieved through the PPS signal provided by the Beidou navigation system;

[0089] 4. Data caching: Set up a double-caching mechanism to ensure the continuity and integrity of data collection.

[0090] S2: Based on the real-time position data provided by the Beidou navigation system and the attitude data provided by the inertial measurement unit, the point cloud data is calibrated in space coordinates;

[0091] In S2, the LiDAR and GNSS systems are started, and the device attitude information is obtained through the IMU to achieve real-time calibration of the LiDAR data. Data collection can be performed on different platforms, and the control system automatically switches to the corresponding operation mode (such as drone high-altitude mode, vehicle-mounted mode or portable mode), and performs data processing in the background. The collected point cloud data is transmitted to the data processing module, and the point cloud data is cleaned, spliced ​​and denoised using a dedicated algorithm, and spatial coordinates are calibrated to generate terrain data that meets application requirements.

[0092] S2 specifically includes the following sub-steps:

[0093] S2.1: Organize the point cloud data collected in different time series and different spatial positions into time and space sequences;

[0094] S2.2: Analyze point cloud data of different time series to extract long-term change trends of landforms;

[0095] S2.3: Based on point cloud data, analyze the geomorphic structural characteristics from different spatial perspectives and explore the three-dimensional morphological change laws of geomorphic structural characteristics;

[0096] S2.4: Preprocess the point cloud data based on the long-term change trend and the three-dimensional morphological change law;

[0097] The detailed description of S2 and its sub-steps (S2.1-S2.4) is as follows:

[0098] S2 mainly involves the spatial coordinate calibration process of point cloud data. In multi-platform measurement scenarios, due to the influence of factors such as measurement environment, platform attitude and sensor error, the original point cloud data has spatial registration errors. Therefore, it is necessary to use the data provided by the Beidou navigation system and inertial measurement unit for accurate calibration.

[0099] Specifically, the BeiDou navigation system uses a multi-mode and multi-frequency receiver, supports joint positioning of multiple systems such as BeiDou-3, GPS and GLONASS, and has a positioning accuracy better than 2cm+1ppm (RMS). The inertial measurement unit uses a fiber optic gyroscope combined navigation system, with a gyroscope zero bias stability better than 0.01° / h, an accelerometer zero bias stability better than 50μg, and an attitude accuracy better than 0.008°.

[0100] Before performing spatial coordinate calibration, the collected point cloud data must first be organized and preprocessed in time and space sequence (S2.1-S2.4):

[0101] S2.1 Spatiotemporal sequence organization of point cloud data:

[0102] Establish a unified time base: Use UTC time as the unified time base; align timestamps for data collected on different platforms; establish a mapping relationship between GPS time and system time;

[0103] Construct spatiotemporal index structure: use octree structure to perform spatial indexing on point cloud data; establish B+ tree index based on timestamp; realize fast retrieval mechanism of spatiotemporal multidimensional index;

[0104] Data shard management: Divide the survey area into regular grids of fixed size (e.g. 100m×100m); assign a unique spatiotemporal identifier to each data shard; establish an adjacency table between shards;

[0105] S2.2 Analysis of long-term landform change trends:

[0106] Comparative analysis of historical data:

[0107] Collect historical surveying and mapping data of the study area; establish the registration relationship of multi-temporal point cloud data; extract key feature points of landform changes;

[0108] Trend modeling:

[0109] Construct a time series change model; analyze the evolution law of terrain relief; establish a landform evolution prediction model;

[0110] Abnormal change detection:

[0111] Set change threshold parameters; identify mutation areas and gradual change areas; record the temporal and spatial distribution patterns of change characteristics;

[0112] Before analyzing the long-term trend of landform change, although the full integration of multi-platform point cloud data has not been completed, the system has completed the organization of spatiotemporal sequences (S2.1). At this stage, the original point cloud data collected by each platform has established a unified time reference, spatiotemporal index structure and data sharding management, which provides a basis for preliminary trend analysis.

[0113] Specifically, in one embodiment, the system adopts a hierarchical and progressive trend analysis strategy. The trend analysis before point cloud fusion is mainly based on the following aspects:

[0114] First, through the spatiotemporal index structure, the system can identify and track the changes of key geomorphic feature points at a coarse-grained level. These feature points are usually stable and easily recognizable landmark structures in the landform, such as the outline of a cave entrance, a significant rock protrusion, etc. These feature points can be reliably identified in the measurement data of each platform, and time series comparison can be performed without the need for complete point cloud fusion.

[0115] Secondly, based on the data sharding management mechanism, the system performs independent statistical analysis on the point cloud data in each grid unit. By comparing the statistical features of the same grid unit at different time points (such as average elevation, point cloud density distribution, etc.), the overall trend of landform changes can be preliminarily grasped. This analysis based on statistical features has low requirements for the precise fusion of point clouds.

[0116] Thirdly, the system uses historical surveying and mapping data to establish a baseline model of landform change. This baseline model contains the general laws and typical patterns of landform evolution, which can guide the system to identify abnormal change trends before point cloud fusion. For example, when the rate of change in certain areas deviates significantly from the historical pattern, the system can issue a warning in time even before complete fusion.

[0117] Finally, the system uses a multi-scale analysis method to perform trend analysis at different spatial scales. Large-scale trend analysis focuses on the overall changes in landforms, which requires relatively low precision fusion of point cloud data; while detailed level change analysis is left to be performed after point cloud fusion. This hierarchical analysis strategy ensures the reliability and efficiency of trend analysis.

[0118] It should be noted that although the accuracy of this early trend analysis may not be as accurate as the fused analysis, it has important early warning value. By discovering potential risk trends early, the system can provide directional guidance for subsequent detailed analysis. After completing the point cloud fusion, the system will verify and refine the early trend analysis results to obtain a more accurate assessment of landform changes.

[0119] This strategy of trend analysis first and data fusion later meets the actual needs of geological disaster early warning. In some cases, timely detection of risk trends is more important than obtaining high-precision changes. By conducting trend analysis in the early stages of data processing, the system can better balance the requirements of real-time and accuracy, and provide more effective technical support for geological disaster prevention.

[0120] S2.3 Analysis of geomorphic structure characteristics:

[0121] Multi-view data fusion: Integrate point cloud data obtained from different platforms; establish a view transformation matrix; eliminate system errors caused by view differences;

[0122] Geomorphic feature extraction: calculate terrain slope and slope aspect; identify characteristic lines and characteristic points; extract morphological features of landforms;

[0123] Analysis of three-dimensional morphological changes: constructing landform descriptors; analyzing the spatial distribution characteristics of terrain fluctuations; establishing mathematical models of landforms;

[0124] S2.4 Point cloud data preprocessing:

[0125] B1. Data cleaning:

[0126] Remove outliers and duplicate points; correct reflection intensity outliers; fill in data hole areas;

[0127] B2. Density homogenization treatment:

[0128] Calculate local point cloud density; interpolate data in sparse areas; downsample overcrowded areas;

[0129] B3. Coordinate system conversion:

[0130] Convert the WGS84 coordinate system to the local coordinate system; apply the seven-parameter conversion model while taking into account the effect of the earth's curvature;

[0131] After the above preprocessing, the spatial coordinates are calibrated using the data from the Beidou navigation system and the inertial measurement unit:

[0132] Attitude data processing: Apply Kalman filter algorithm to process IMU data; eliminate attitude angle drift; calculate the real-time attitude matrix of the platform;

[0133] Position data processing: differential decomposition of Beidou observation data; multi-system combined positioning; centimeter-level positioning accuracy;

[0134] Spatiotemporal registration: establish the spatial relationship between the sensor and the carrier; calculate the absolute coordinates of the point cloud data; achieve accurate synchronization of multi-source data.

[0135] The system analyzes the changes in data collected in different time periods to identify the long-term evolution trend and short-term fluctuation characteristics of the landform. Through this time series exploration, we can more comprehensively understand the dynamic changes of complex structures such as caves and rocks in karst landforms, and further improve the accuracy of data analysis.

[0136] In addition, karst landforms usually have dynamically changing three-dimensional structures due to their complex geological features (such as karst, caves, cliffs, etc.), and their topography evolves over time and with changes in the environment. Traditional single-shot mapping methods can only capture static data, making it difficult to fully understand the diversity and changing trends of karst landforms. In order to better understand these complex landforms and their changes, sequence exploration technology can be used to analyze mapping data at different time points and spaces to explore the dynamic changes of landform features.

[0137] Implementation process

[0138] Step C1: Apply sequence exploration to mine landform features

[0139] The system obtains a large amount of LiDAR point cloud data from multiple surveying and mapping tasks, which may be collected at different times and different spatial locations, forming a continuous spatiotemporal data sequence. For example, surveying and mapping tasks may be carried out in different seasons of the year or at multiple locations in the same area, thus capturing the characteristics of karst landforms that change over time and environment.

[0140] The system organizes these point cloud data from different times and spaces into a spatiotemporal sequence. For example, caves in karst areas may experience rock collapse or water erosion as the seasons change. After arranging these sequence data in chronological order, the system can analyze the evolution of caves in spring, summer, and winter, or observe the changes in karst characteristics in different regions.

[0141] The system gradually extracts the potential features of the landform from these spatiotemporal sequence data through a sequence exploration algorithm. The sequence exploration algorithm can discover changes in karst landforms through frame-by-frame analysis, such as the long-term deposition of certain rock structures and the dynamic process of crack expansion.

[0142] Local feature mining: By comparing mapping data from adjacent time points, the system can mine small-scale changes, such as early signs of rock collapse or cave expansion.

[0143] Overall feature mining: By conducting spatiotemporal comprehensive analysis of data from the entire terrain area, the system can capture changes in large structures, such as surface collapse and the formation of karst depressions.

[0144] Sequence exploration can not only analyze landform changes from a temporal dimension, but also perform feature mining from different spatial perspectives. For example, the system can analyze the structure of a karst cave from the top, side, or bottom perspectives, and gradually dig out the details of its three-dimensional morphological changes. This multi-dimensional exploration method can provide richer landform information than a single perspective, especially when analyzing complex karst structures or deep underground caves.

[0145] Step C2: Analysis of long-term and short-term patterns of sequence data

[0146] By comparing point cloud data collected at different time periods, the system can identify long-term trends in karst landforms. For example, analyzing data from many years can identify the chronic expansion of certain karst caves or the long-term erosion of certain rock layers.

[0147] Cave expansion trends: When analyzing sequential data for caves, the system may find a trend in cave entrances expanding year by year. This long-term trend may reflect the impact of groundwater flow on rock structure, or increased rock erosion due to climate change.

[0148] Evolution of karst depressions: By comparing surveying and mapping data over the years, the system may identify that karst depressions in certain areas are continuing to expand, gradually forming more obvious topographic changes. This long-term change trend can provide an important basis for the prediction of geological disasters.

[0149] In addition to long-term trends, sequence exploration can also detect short-term changes and fluctuations in karst landforms. For example, after seasonal rainfall or short-term fluctuations in groundwater levels, the system can identify short-term collapse, crack expansion, or groundwater flushing in caves.

[0150] Seasonal changes: The system can identify short-term collapse or crack expansion of certain rocks during the rainy season by analyzing seasonal series data. Such short-term fluctuations are very important for the prevention and monitoring of geological disasters.

[0151] Short-term erosion phenomena: By analyzing data collected repeatedly over a short period of time, the system can detect rapid erosion of rock surfaces caused by short-term water flow or weathering.

[0152] The system will show the long-term trends and short-term fluctuations of landforms through dynamic visualization. For example, through animation or multidimensional models, the system can dynamically show how karst landforms change in shape over months or years. This visualization method can not only intuitively reflect the process of landform evolution, but also help decision makers better understand geological risks.

[0153] After each analysis is completed, the system will feed back the results of the sequence exploration into the data model library to optimize future predictions and analysis. This feedback mechanism can help the system gradually accumulate experience and more quickly and accurately identify dynamic changes in landforms in similar surveying and mapping tasks in the future.

[0154] Example

[0155] Consider a five-year geomorphic survey in a karst region. Each year, the survey task collected a large amount of point cloud data, focusing on the changes in a large cave.

[0156] M1. Long-term trend analysis

[0157] The system analyzed five years of sequence data and found that the entrance of the cave expanded by 2-5 meters every year, and the rate of rock erosion was accelerating year by year. This trend showed that the cave might undergo structural collapse in the next few years, and the system generated a geological warning report based on this.

[0158] M2. Short-term volatility analysis

[0159] During the rainy season of the year, mapping data showed that cracks expanded in some areas of the cave in a short period of time. By analyzing the data series each month, the system can accurately identify that these cracks gradually expanded after several heavy rains and issue short-term geological disaster risk warnings.

[0160] M3. Multi-perspective exploration

[0161] By analyzing the cave data from different spatial perspectives (such as from the top, side, etc.), the system found that the rock layer at the top of the cave was thinning, which further confirmed that the cave had potential risk of collapse.

[0162] Summarize

[0163] Through sequence exploration technology, the system can not only analyze the long-term change trend of karst landforms from a temporal dimension, but also capture short-term fluctuation characteristics, thereby providing a more accurate and comprehensive understanding of complex terrain. Sequence exploration technology combines time and space sequence data to significantly improve the ability to mine landform data and help better conduct geological risk assessment and landform evolution research.

[0164] S3: input the calibrated point cloud data into the planning-aware diffusion network for first noise removal;

[0165] The complexity of karst landforms (such as irregular rock structures, deep caves, steep mountains, etc.) can cause noise problems in lidar data, making it difficult to accurately complete traditional point cloud cleaning and prediction. To solve this problem, the planning-aware diffusion network can be combined with technologies with offline value function memory and sequence exploration to accurately predict and optimize these landform features. This method enhances the removal of point cloud noise by learning and exploring sequence data through the network, further improving the data processing accuracy and the quality of landform reconstruction.

[0166] S3.1: construct terrain data segments based on point cloud data, and input the terrain data segments into the planning-aware diffusion network;

[0167] From the point cloud data collected by the LiDAR, data fragments of karst landform features are extracted and input into the planning-aware diffusion network. This network can effectively perceive and identify noise points in landform data by combining deep learning methods with prior planning information.

[0168] S3.2: Predict the first noise point in the local area of ​​the terrain data segment through the planning-aware diffusion network;

[0169] The diffusion network will gradually predict the noise points in the local area based on the current input terrain data fragment, and determine which points are noise points based on the feature propagation in the diffusion process. The advantage of this network is that it can perceive the subtle differences in complex terrain and effectively identify noise points under complex shapes.

[0170] S3.3: Compare the predicted first noise point information with the landform planning to verify the consistency between the first noise point information and the actual terrain features;

[0171] After each local prediction is completed, the system will compare the noise information with the landform planning through a feedback mechanism to ensure that the identified noise points are consistent with the actual terrain features. Through multiple rounds of iterations, the prediction accuracy of the diffusion network is gradually optimized to adapt it to the complexity of karst landforms.

[0172] S3.4: After the verification is completed, the first noise point in the calibrated point cloud data is removed;

[0173] The planning perception diffusion network is the core technical module used for the first noise removal in this application. Figure 2 ,The network adopts an encoder-decoder structure, and uses a multi-layer perceptron and attention mechanism to ,accurately identify and remove noise points in point cloud data.,Before noise processing, it is necessary to ,first construct terrain data segments, which are divided based on the ,spatial distribution characteristics of point cloud data, and each data segment contains the ,terrestrial feature information of the local area.

[0174] In the process of constructing data segments, the system uses an adaptive segmentation algorithm to dynamically adjust the segment size according to the point cloud density and terrain complexity. For areas with drastic terrain changes, the segment size is smaller, usually 10m×10m; while for areas with gentle terrain, the segment size can be appropriately increased, up to 50m×50m. In addition to spatial coordinate information, each data segment also contains characteristic attributes such as reflection intensity and number of echoes.

[0175] After receiving the terrain data fragment, the planning perceptual diffusion network first extracts local features through a multi-layer perceptron. The perceptron adopts a residual structure and contains multiple convolutional layers and pooling layers, which can effectively extract the geometric and semantic features of point cloud data. The self-attention mechanism is introduced in the middle layer of the network to enhance the understanding of spatial structure by calculating the correlation between points. Based on the feature extraction, the network uses a diffusion model to predict the first noise point in the local area.

[0176] During the prediction process, the diffusion model learns the potential distribution of point cloud data by gradually adding and removing noise. The model sets 1,000 diffusion steps. In each step, the data distribution at the next moment is predicted based on the data state and noise level at the current moment. This diffusion-based prediction method is highly robust and can adapt to different types of noise distribution.

[0177] In order to ensure the accuracy of noise point prediction, the system compares the prediction results with the landform planning. The landform planning information contains key information such as the terrain characteristics, building distribution, and vegetation coverage of the survey area. The comparison process uses a multi-scale matching algorithm to verify the rationality of the prediction results by calculating the spatial correlation between the predicted noise points and landform features. If there is an obvious conflict between the predicted noise points and the landform features, the system will correct the prediction results.

[0178] The following is a description of an example case scenario:

[0179] Suppose a LiDAR is used to survey a karst area with complex terrain, including caves, underground cracks, and irregular limestone landforms. Since the light emitted by the LiDAR may be affected by complex surface reflection and scattering, the point cloud data in the measurement results often contains noise (such as erroneous reflections from the edge of the rock, echo interference, etc.). These noises will reduce the accuracy of the subsequent 3D model and need to be effectively identified and removed by some algorithm.

[0180] Specific implementation process

[0181] 1). Data collection

[0182] Using drones or vehicle-mounted LiDAR equipment to survey a karst area generates point cloud data containing a large amount of terrain information. This data may contain noise, such as reflection errors at the edge of a cliff or noise interference in underground cracks.

[0183] 2) Application of planning perception diffusion network

[0184] Network initialization

[0185] The system uses a pre-trained planning-aware diffusion network to process point cloud data. This network has been trained with a large amount of prior data on karst landforms and can identify which terrain features are real and which are noise points.

[0186] Noise Identification Process

[0187] After receiving new LiDAR data, the network first predicts each point based on the data's geographic location information and the planned landform features. For karst landforms, points in caves and crack structures are prone to anomalies. The network will gradually perceive and identify those points that do not match the surrounding terrain features through a diffusion process of adjacent points.

[0188] For example, at the edge of a cave, a laser may mistakenly detect distant reflections that appear to be related to the surrounding terrain, but because of their unusual positions or angles, the diffusion network can identify them as noise and mark them out.

[0189] Noise prediction mechanism

[0190] The diffusion network gradually analyzes the local area of ​​the point cloud data through the diffusion strategy. It perceives the position relationship of each point in the overall terrain and determines whether the point meets the natural characteristics of karst landforms. For example:

[0191] If a point deviates significantly from the natural curve of the rock wall, the diffusion network predicts it to be a noise point based on the distribution of its neighboring points.

[0192] The network will also identify some points as noise if their distribution spatially does not match typical karst landforms (e.g., randomly distributed point clouds instead of continuous rock surfaces).

[0193] 3). Noise removal and feedback

[0194] Noise Removal

[0195] The system marks those point clouds predicted to be noise points under the processing of the diffusion network. Then, these noise points are removed from the point cloud data through the automatic cleaning function.

[0196] Feedback Optimization

[0197] After each noise removal, the system generates a feedback mechanism to compare the processed data with the original terrain plan to ensure that the removed noise does not affect the accuracy of the real terrain. If some noise points are found to be misidentified, the system will fine-tune the diffusion network through feedback to further improve the accuracy of recognition.

[0198] Example Results

[0199] After processing the point cloud data of this karst landform area, the original point cloud map containing noise has been optimized, and the 3D model of complex terrain such as the edge of the cave and the depth of the cracks is more accurate. By planning the perception diffusion network, the system successfully removed the point cloud interference caused by false reflections, making the model more accurate and reliable.

[0200] The embodiment of the present invention also includes:

[0201] D1: Build a value function library based on historical surveying and mapping data and landform models, which stores point cloud noise processing strategies under different terrain scenarios;

[0202] The system uses a large amount of offline data (including historical surveying and mapping data, landform models, etc.) to build a value function library, which records the best strategies for optimizing point cloud noise processing in different scenarios. This process is similar to the reward mechanism in reinforcement learning, which aims to quickly identify which strategies can effectively reduce redundant work in data processing.

[0203] D2: According to the characteristics of the current point cloud data, retrieve and call the best point cloud noise processing strategy from the value function library;

[0204] D3: Use the best point cloud noise processing strategy to process the first noise point in the current point cloud data;

[0205] In practical applications, in order to improve the efficiency and accuracy of noise processing, this application constructs a value function library based on historical surveying and mapping data. The function library contains typical noise distribution patterns and processing strategies in different terrain scenarios. For example, in karst cave areas, due to the influence of laser reflection and multipath effects, specific patterns of noise are often generated. The value function library stores the optimal processing strategies for these typical scenarios.

[0206] When the system processes new point cloud data, it first analyzes the characteristic attributes of the current data, including point cloud density distribution, elevation change characteristics, reflection intensity distribution, etc. Then, through the similarity matching algorithm, it retrieves the most similar scene mode from the value function library and calls the corresponding noise processing strategy. This experience-based strategy selection method can significantly improve the efficiency of noise processing.

[0207] The execution process of the optimal point cloud noise processing strategy is adaptive. The system will dynamically adjust the processing parameters, such as the noise determination threshold, spatial clustering radius, etc., according to the characteristics of the current point cloud data. During the processing, the terrain characteristics of the local area will also be considered, and the processing intensity will be finely adjusted to ensure that the effective terrain information will not be destroyed while removing the noise.

[0208] After processing, the system will conduct a quality assessment of the results, including integrity check of point cloud data, local density analysis, and noise residual detection. If the processing result is found to be unsatisfactory, the system will automatically adjust the processing strategy and perform secondary optimization. Finally, the point cloud data after noise processing will be used in the subsequent classification and modeling process.

[0209] The following is a description of a practical scenario:

[0210] Assume that the present embodiment of the application performs LiDAR mapping in a karst landform area and generates a large amount of point cloud data. Due to the complexity of the landform, such as caves, cracks, cliffs and other features, there will be a lot of noise in the data. In order to improve data processing efficiency, the present embodiment of the application uses the "offline value function memory" technology to quickly optimize the noise recognition and processing in the current mapping task through the experience in historical mapping data and landform models.

[0211] Specific implementation process

[0212] E1: Building a value function library

[0213] The system has collected a large amount of historical surveying and mapping data on karst landforms. These data come from different operating scenarios, such as mountainous areas, forests, underground caves, etc., and cover surveying and mapping tasks under various complex terrains. These historical data include strategies and effects for processing noise.

[0214] The system uses these historical data to build multiple three-dimensional models of karst landforms, including the shape of rock caves, the direction of cracks, the depth of underground cavities, etc. These models record how to deal with noise and which strategies are most effective.

[0215] Based on the performance of different strategies in historical mapping tasks, the system defines a "value function" for noise processing in each scene. This function is similar to the reward mechanism in reinforcement learning. When a processing strategy successfully reduces noise without losing data accuracy, the "value" of the strategy will increase. For example, when processing cliff edge data, if a specific algorithm can effectively remove noise from false reflections, the strategy will be given a higher value. Through such training, the system gradually established a complete value function library, covering the best strategies for optimizing noise processing under various terrain conditions.

[0216] E2: Calling offline value functions for real-time optimization

[0217] In the new karst landform mapping task, the system received point cloud data collected by drones. These data contain complex terrain such as rocks, cracks, cliffs, etc., and also contain noise.

[0218] Based on the terrain features of the current mapping task, the system automatically queries the value function library to find the historical data processing strategy that best matches these terrains. For example, if the system recognizes that the current mapping data is similar to a certain valley scene in history, it will immediately call the high-value strategy related to the scene, such as using a special filtering algorithm to remove the false reflection point cloud at the edge of the cave.

[0219] Real-time optimization of processing workflows, including:

[0220] Strategy 1: Noise removal based on historical cliff data

[0221] At the edge of a cliff, the LiDAR may capture reflections from far away, generating noise. Based on the records in the value function library, the system calls the best noise removal strategy for cliff terrain in the past, which has successfully processed false point clouds in similar terrain. The system immediately applies this strategy and quickly removes interference points in the data.

[0222] Strategy 2: Dealing with noise in underground caves

[0223] In underground caves, LiDAR may generate noise due to abnormal light reflection. The system calls another strategy in the value function library, which comes from past experience in dealing with complex underground terrain. Through spatial correction and filtering algorithms for point cloud data, the system quickly removes false reflection data in caves.

[0224] Because the system calls the best historical strategy from the value function library, it avoids recalculating the noise processing algorithm for each new scene, greatly speeding up the efficiency of data processing. For example, the noise that originally required a lot of repeated calculations to identify and process can be completed with only a small amount of calculation after calling the value function, and the processing speed is increased several times.

[0225] After the processing is completed, the system will feed back the strategy and processing results used in the current scenario to the value function library, and update the value function based on the processing effect. If the processing strategy in the new scenario performs well, the system will add it to the value library so that it can be called in similar scenarios in the future.

[0226] Example Results

[0227] Through the "offline value function memory" technology, the system quickly calls the best noise processing strategy when processing karst landform data, reducing repeated calculations and improving processing efficiency. In actual applications, the system uses historical high-value strategies to achieve accurate noise removal when processing complex terrain such as cliff edges and underground caves, and ensures the overall accuracy and consistency of the data. Data processing time is shortened by 30% to 50%.

[0228] Here is the sample code:

[0229] / / Value function library class definition

[0230] class ValueFunctionLibrary {

[0231] constructor() {

[0232] / / Initialize the parameters and historical data of the value function library

[0233] this.historicalData = new Map();

[0234] this.strategies = new Map();

[0235] this.valueThresholds = {

[0236] minPointDensity: 5, / / minimum point cloud density (points / square meter)

[0237] maxIsolation: 0.8, / / Maximum isolation threshold

[0238] minReflectivity: 0.1, / / minimum reflection intensity

[0239] maxReflectivity: 0.9 / / Maximum reflection intensity

[0240] };

[0241] }

[0242] / / Add historical surveying data

[0243] addHistoricalData(scenarioId, data) {

[0244] const processedData = this._preprocessHistoricalData(data);

[0245] this.historicalData.set(scenarioId, processedData);

[0246] this._updateStrategies(scenarioId);

[0247] }

[0248] / / Preprocess historical data

[0249] _preprocessHistoricalData(data) {

[0250] return {

[0251] pointDensity: this._calculatePointDensity(data.points),

[0252] terrainType: this._classifyTerrainType(data.elevation),

[0253] noiseLevel: this._evaluateNoiseLevel(data.points),

[0254] processingEffectiveness: data.processingEffectiveness || 0.8

[0255] };

[0256] }

[0257] / / Calculate point cloud density

[0258] _calculatePointDensity(points) {

[0259] / / Calculate the point cloud density in a given area

[0260] const area = this._calculateArea(points);

[0261] return points.length / area;

[0262] }

[0263] / / Terrain classification

[0264] _classifyTerrainType(elevation) {

[0265] / / Classify terrain based on elevation data

[0266] const slope = this._calculateSlope(elevation);

[0267] const roughness = this._calculateRoughness(elevation);

[0268] if (slope>45&&roughness>0.7) {

[0269] return 'steep_karst';

[0270] } else if (slope>30&&roughness>0.5) {

[0271] return 'moderate_karst';

[0272] } else {

[0273] return 'gentle_karst';

[0274] }

[0275] }

[0276] / / Evaluate the noise level

[0277] _evaluateNoiseLevel(points) {

[0278] let isolatedPoints = 0;

[0279] let totalPoints = points.length;

[0280] / / Calculate the proportion of isolated points

[0281] for (const point of points) {

[0282] if (this._isIsolatedPoint(point, points)) {

[0283] isolatedPoints++;

[0284] }

[0285] }

[0286] return isolatedPoints / totalPoints;

[0287] }

[0288] / / Update processing strategy

[0289] _updateStrategies(scenarioId) {

[0290] const data = this.historicalData.get(scenarioId);

[0291] if (!data) return;

[0292] / / Develop processing strategies based on historical data

[0293] const strategy = {

[0294] filteringThreshold: this._calculateFilteringThreshold(data),

[0295] segmentationParams: this._determineSegmentationParams(data),

[0296] noiseRemovalMethod: this._selectNoiseRemovalMethod(data),

[0297] valueScore: this._calculateValueScore(data)

[0298] };

[0299] this.strategies.set(scenarioId, strategy);

[0300] }

[0301] / / Calculate the filtering threshold

[0302] _calculateFilteringThreshold(data) {

[0303] const baseThreshold = 0.5;

[0304] const densityFactor = Math.min(1, data.pointDensity / this.valueThresholds.minPointDensity);

[0305] const noiseFactor = 1 - data.noiseLevel;

[0306] return baseThreshold * densityFactor * noiseFactor;

[0307] }

[0308] / / Determine the segmentation parameters

[0309] _determineSegmentationParams(data) {

[0310] return {

[0311] minSegmentSize: Math.max(10, Math.floor(data.pointDensity * 2)),

[0312] maxSegmentSize: Math.min(1000, Math.floor(data.pointDensity * 20)),

[0313] growthThreshold: 0.1 + (data.noiseLevel * 0.2)

[0314] };

[0315] }

[0316] / / Select the noise removal method

[0317] _selectNoiseRemovalMethod(data) {

[0318] if (data.noiseLevel>0.6) {

[0319] return'statistical_outlier_removal';

[0320] } else if (data.pointDensity<this.valueThresholds.minPointDensity) {

[0321] return 'radius_outlier_removal';

[0322] } else {

[0323] return 'conditional_removal';

[0324] }

[0325] }

[0326] / / Calculate the value score

[0327] _calculateValueScore(data) {

[0328] const densityScore = Math.min(1, data.pointDensity / this.valueThresholds.minPointDensity);

[0329] const noiseScore = 1 - data.noiseLevel;

[0330] const effectivenessScore = data.processingEffectiveness;

[0331] return (densityScore * 0.3 + noiseScore * 0.4 + effectivenessScore *0.3);

[0332] }

[0333] / / Get the best processing strategy

[0334] getBestStrategy(currentData) {

[0335] const currentFeatures = this._extractFeatures(currentData);

[0336] let bestStrategy = null;

[0337] let highestSimilarity = -1;

[0338] / / Traverse historical strategies to find the most similar scenario

[0339] for (const [scenarioId, strategy] of this.strategies) {

[0340] const similarity = this._calculateSimilarity(

[0341] currentFeatures,

[0342] this.historicalData.get(scenarioId) );

[0344] if (similarity>highestSimilarity) {

[0345] highestSimilarity = similarity;

[0346] bestStrategy = strategy;

[0347] }

[0348] }

[0349] return bestStrategy;

[0350] }

[0351] / / Extract features

[0352] _extractFeatures(data) {

[0353] return {

[0354] pointDensity: this._calculatePointDensity(data.points),

[0355] terrainType: this._classifyTerrainType(data.elevation),

[0356] noiseLevel: this._evaluateNoiseLevel(data.points)

[0357] };

[0358] }

[0359] / / Calculate similarity

[0360] _calculateSimilarity(features1, features2) {

[0361] const densityDiff = Math.abs(features1.pointDensity - features2.pointDensity)

[0362] / Math.max(features1.pointDensity, features2.pointDensity);

[0363] const terrainMatch = features1.terrainType === features2.terrainType? 1 : 0;

[0364] const noiseDiff = Math.abs(features1.noiseLevel - features2.noiseLevel);

[0365] return (1 - densityDiff) * 0.3 + terrainMatch * 0.4 + (1 - noiseDiff) * 0.3;

[0366] }

[0367] / / Helper method: Calculate area

[0368] _calculateArea(points) {

[0369] / / Simplified area calculation. More complex convex hull algorithms are required in practical applications.

[0370] const xCoords = points.map(p =>px);

[0371] const yCoords = points.map(p =>py);

[0372] const width = Math.max(...xCoords) - Math.min(...xCoords);

[0373] const height = Math.max(...yCoords) - Math.min(...yCoords);

[0374] return width * height;

[0375] }

[0376] / / Helper method: calculate slope

[0377] _calculateSlope(elevation) {

[0378] / / Simplified slope calculation

[0379] const maxElevation = Math.max(...elevation);

[0380] const minElevation = Math.min(...elevation);

[0381] const horizontalDistance = Math.sqrt(elevation.length); / / Assume that the points are evenly distributed

[0382] return Math.atan((maxElevation - minElevation) / horizontalDistance)* (180 / Math.PI);

[0383] }

[0384] / / Helper method: calculate roughness

[0385] _calculateRoughness(elevation) {

[0386] const mean = elevation.reduce((a, b) =>a + b) / elevation.length;

[0387] const variance = elevation.reduce((a, b) =>a + Math.pow(b - mean, 2),0) / elevation.length;

[0388] return Math.sqrt(variance) / (Math.max(...elevation) - Math.min(...elevation));

[0389] }

[0390] / / Helper method: determine whether it is an isolated point

[0391] _isIsolatedPoint(point, points) {

[0392] const threshold = 0.5; / / Isolated point determination threshold

[0393] let nearbyPoints = 0;

[0394] for (const other of points) {

[0395] if (point === other) continue;

[0396] const distance = Math.sqrt(

[0397] Math.pow(point.x - other.x, 2) +

[0398] Math.pow(point.y - other.y, 2) +

[0399] Math.pow(point.z - other.z, 2) );

[0401] if (distance <threshold) nearbyPoints++;

[0402] }

[0403] return nearbyPoints<3; / / If there are less than 3 nearby points, it is considered an isolated point

[0404] }

[0405] }

[0406] / / Example

[0407] function demonstrateValueFunctionLibrary() {

[0408] const library = new ValueFunctionLibrary();

[0409] / / Add historical data example

[0410] const historicalScenario = {

[0411] points: [

[0412] { x: 0, y: 0, z: 0},

[0413] { x: 1, y: 1, z: 1},

[0414] / / ... more data

[0415] ],

[0416] elevation: [0, 1, 2, 1.5, 0.8],

[0417] processingEffectiveness: 0.85

[0418] };

[0419] library.addHistoricalData('scenario1', historicalScenario);

[0420] / / Get the best processing strategy for the current data

[0421] const currentData = {

[0422] points: [

[0423] { x: 0.1, y: 0.1, z: 0.1},

[0424] { x: 1.1, y: 1.1, z: 1.1},

[0425] / / ... more data

[0426] ],

[0427] elevation: [0.1, 1.1, 2.1, 1.6, 0.9]

[0428] };

[0429] const bestStrategy = library.getBestStrategy(currentData);

[0430] console.log('Best strategy:', bestStrategy);

[0431] }

[0432] / / Export the value function library class for use by other modules

[0433] export default ValueFunctionLibrary;

[0434] This value function library implements the following core functions:

[0435] 1. Data Management:

[0436] Maintain historical surveying and mapping data and corresponding processing strategies;

[0437] Support adding new historical data and updating strategies;

[0438] Provide data preprocessing and feature extraction functions;

[0439] 2. Value Assessment:

[0440] Calculate point cloud density, terrain classification, and noise level;

[0441] Comprehensively evaluate data quality based on multiple factors;

[0442] Dynamically adjust processing parameters and thresholds;

[0443] 3. Strategy selection:

[0444] Choose the most appropriate processing strategy for different scenarios;

[0445] Match the best historical experience through similarity calculation;

[0446] Support adaptive parameter adjustment;

[0447] Key Features:

[0448] 1. Terrain classification: Karst landforms are divided into three categories according to slope and roughness: steep_karst (steep), moderate_karst (moderate) and gentle_karst (gentle).

[0449] 2. Noise assessment: The noise level is assessed by analyzing the spatial distribution characteristics of the point cloud and calculating the proportion of isolated points.

[0450] 3. Processing strategy: including complete configuration of filtering thresholds, segmentation parameters, and noise removal methods.

[0451] Directions:

[0452] javascript

[0453] const library = new ValueFunctionLibrary();

[0454] / / Add historical experience data

[0455] library.addHistoricalData('scenario1', historicalData);

[0456] / / Get the best processing strategy for the current data

[0457] const bestStrategy = library.getBestStrategy(currentData);

[0458] S4: using the secondary cutting plane method to classify the point cloud data, dividing the point cloud data into ground point cloud and non-ground point cloud;

[0459] Karst landforms are known for their complex topographic features, with irregular karst landforms, caves, cliffs, and deep fissures. During the surveying process, it is crucial to distinguish between ground point clouds and non-ground point clouds (such as noise caused by trees, buildings, and bad weather). At the same time, bad weather (such as rain, snow, fog, etc.) may introduce noise in non-ground point clouds, affecting the surveying accuracy. The use of "sub-quadratic cut planes" combined with near-optimal deterministic network decomposition and rule sets and the improved MIS (maximum independent set) algorithm can effectively identify noise in bad weather and provide high-precision landform data, which is particularly suitable for complex surveying and mapping scenarios of karst landforms.

[0460] The system obtains point cloud data of karst landforms from the lidar mapping process. The point cloud contains both real features of the ground (such as karst surface and cliffs) and non-ground point clouds (such as trees, buildings, and noise generated by bad weather).

[0461] S4.1: Construct multiple cutting planes based on the height information of the point cloud data and the spatial relationship between adjacent points;

[0462] like Figure 3 As shown, the secondary quadratic cutting plane method is an algorithm that uses geometric segmentation and can effectively distinguish between ground point clouds and non-ground point clouds based on the height information of the point cloud and the spatial relationship between adjacent points.

[0463] The system analyzes the height changes of the terrain, creates multiple cutting planes, and divides the point cloud into two categories: ground (such as rocks and soil) and non-ground (such as trees, buildings, and meteorological noise). For example, in karst landforms, the system can identify terrain such as karst depressions and caves and divide them into ground point clouds.

[0464] S4.2: Marking a point cloud that is closely connected to the ground and has a height below a first threshold as a ground point cloud;

[0465] S4.3: Mark the point cloud separated from the ground or with a height higher than a first threshold as a non-ground point cloud;

[0466] The cutting plane marks the point clouds with low height and close connection to the ground as "ground point clouds", such as the limestone surface in karst landforms and the outer wall of caves.

[0467] Point clouds that are high, separated from the ground, or have irregular distributions are marked as "non-ground point clouds", such as noise generated by tree canopies, buildings, or bad weather.

[0468] S4.4: Store ground point cloud and non-ground point cloud separately;

[0469] In the above S4.1-S4.4, after completing the noise processing, the point cloud data needs to be classified into ground point cloud and non-ground point cloud. This application adopts the secondary cutting plane method for classification processing, which can effectively process the complex terrain structure under karst landform while maintaining high classification accuracy.

[0470] The core idea of ​​the secondary cutting plane method is to gradually separate the ground point cloud and the non-ground point cloud by constructing multiple local cutting planes. In the specific implementation process, it is first necessary to construct the cutting plane based on the height information of the point cloud data and the spatial relationship between adjacent points. The system adopts an adaptive grid division strategy to divide the survey area into several grid cells. The size of each grid cell is dynamically adjusted according to the complexity of the terrain, usually between 5-20 meters. In each grid cell, the initial estimate of the local ground is obtained by least squares fitting.

[0471] In order to improve the accuracy of cutting plane construction, the system introduces a terrain gross error detection mechanism. By calculating the distance from each point to the local plane and combining statistical test methods, outliers that are obviously deviated from the local ground are identified and removed. On this basis, the iterative weighted least squares algorithm is used to optimize the cutting plane parameters. In each iteration, the weight is dynamically adjusted according to the distance from the point to the plane, so that the cutting plane better fits the real terrain.

[0472] After the cutting plane is constructed, the system starts to classify the point cloud. For each point, the distance to the adjacent cutting plane and the normal vector angle are calculated first. If the distance from the point to the cutting plane is less than the preset first threshold (the threshold is dynamically adjusted according to the measurement accuracy and terrain characteristics, usually between 0.2-0.5 meters), and the normal vector angle meets certain conditions, the point is marked as a ground point cloud. The normal vector angle threshold here is usually set between 15-30 degrees, which can be adjusted according to the degree of terrain undulation.

[0473] For point clouds that are separated from the ground or whose height exceeds the first threshold, the system marks them as non-ground point clouds. In this process, the continuity characteristics of the point cloud also need to be considered. If the height of a point exceeds the threshold, but the distribution of its surrounding points indicates that this may be a sudden change in the terrain (such as a steep cliff), the system will decide whether to classify it as a ground point cloud through local connectivity analysis. This classification method that takes spatial continuity into account can better handle special terrains such as steep cliffs and sinkholes that are common in karst landforms.

[0474] In order to improve the reliability of classification, the system also introduces a multi-scale analysis mechanism. By classifying at different scales and fusing the results, the classification error caused by a single scale can be effectively reduced. In the fusion process, a confidence-based weighting strategy is adopted to integrate the classification results of different scales. For areas with inconsistent classification results, the system will conduct key analysis and introduce manual auxiliary judgment when necessary.

[0475] After the classification is completed, the system will store the ground point cloud and non-ground point cloud separately. The storage format adopts the LAS1.4 standard and supports the recording of point cloud classification labels and custom attributes. At the same time, the system will also generate a classification statistical report, including statistical information such as the number distribution and density distribution of ground point clouds and non-ground point clouds, which can be used for subsequent quality assessment and analysis.

[0476] It is particularly important to point out that when processing non-ground point clouds, the system will record the classification basis and credibility of each point. This information is of great reference value for subsequent sub-classification of non-ground point clouds (such as distinguishing vegetation, buildings, etc.). At the same time, the system has also established a version management mechanism for classification results to support the backtracking and correction of classification results.

[0477] In addition, the embodiment of the present application also includes:

[0478] Constructing a connection relationship graph of the non-ground point cloud;

[0479] Decomposing the nodes of the connection relationship graph using a deterministic network decomposition algorithm;

[0480] identifying different types of features in the non-ground point cloud based on height distribution, density, and spatial relationships;

[0481] analyzing tree and building features in the non-ground point cloud using a preset rule set;

[0482] The second noise point in the non-ground point cloud is processed using a maximum independent set algorithm.

[0483] Specifically, the deterministic network decomposition:

[0484] To further refine the classification of non-ground point clouds, the system uses a near-optimal deterministic network decomposition to divide the non-ground point clouds into multiple subsets. These subsets may include different types of features, such as trees, buildings, raindrops, fog and other noise points.

[0485] The system builds a connection graph between point clouds and uses a deterministic algorithm to decompose the nodes in the graph to identify different feature categories. For example, a point cloud with a high degree of stable distribution and clustering may belong to a building or a tree, while a point cloud with a scattered and irregular distribution may be a noise point.

[0486] Applying rule sets for feature recognition includes:

[0487] Rule sets are used to identify different object types in non-ground point clouds. The rule sets analyze features such as height, density, and spatial relationships of the point clouds. For example, the system identifies the top structure of a tree based on the height distribution rules and distinguishes it from possible meteorological noise.

[0488] Some prominent features in karst landforms, such as vegetation or buildings above caves, can be accurately identified using these rule sets, while noise points in severe weather are individually labeled based on their randomly distributed characteristics.

[0489] The second noise point in the non-ground point cloud is processed using a maximum independent set algorithm, including:

[0490] Calculate the spatial independence parameters of point clouds and identify point clouds with discontinuous distribution;

[0491] Analyze the reflection intensity and isolation degree of the point cloud to determine the discrimination threshold of the second noise point;

[0492] Identifying and removing the second noise point based on the discrimination threshold;

[0493] Performing integrity assessment on the non-ground point cloud data after removing the second noise point;

[0494] The discrimination threshold of the second noise point is dynamically adjusted according to the integrity evaluation result.

[0495] When processing the second noise points of non-ground point clouds, the system adopts a noise point identification and processing method based on the maximum independent set theory. The core idea of ​​this method is to construct the point cloud data into an undirected graph, and identify and remove abnormal points by analyzing the spatial relationship and attribute characteristics between points.

[0496] First, the system needs to calculate the spatial independence parameters of the point cloud. This process includes the following steps:

[0497] F1. Establish spatial index structure: Use octree to divide point cloud data into spaces to improve the efficiency of subsequent nearest neighbor search. For each point, the system sets a search radius R (this radius is usually adaptively determined based on the point cloud density, generally 3-5 times the average point spacing).

[0498] F2. Calculate local features: For each point P, count the following features within its search radius R:

[0499] The number of neighboring points N;

[0500] Local point cloud density D = N / sphere volume;

[0501] The average distance to the nearest neighbor is μ d ;

[0502] The standard deviation of the distance σ d ;

[0503] These features together constitute a spatially independent descriptor of point P.

[0504] F3. Construct point cloud connection relationship: If the distance between two points is less than the threshold (usually 2μ d ), then it is considered that the two points are connected. Based on this connection relationship, the adjacency matrix of the point cloud can be constructed.

[0505] After obtaining the spatial independence parameters, the system begins to analyze the reflection intensity and isolation of the point cloud. This stage focuses on:

[0506] H1. Reflection intensity analysis:

[0507] Calculate the reflection intensity histogram of the entire point cloud, use the Gaussian mixture model to model the reflection intensity distribution, and identify abnormal reflection intensity values, which are usually caused by measurement errors or other interference factors;

[0508] H2. Assessment of isolation level:

[0509] Define the isolation index ISO = (σ d / μ d ) * (D max / D);

[0510] Among them, D max is the local maximum density;

[0511] A higher ISO value indicates that the point is relatively isolated;

[0512] The initial threshold T0 of isolation degree is determined by statistical analysis;

[0513] Based on the above analysis, the system uses an adaptive method to determine the discrimination threshold of the second noise point:

[0514] R1. Initial threshold setting:

[0515] T_iso = mean(ISO) + 2 * std(ISO) / / isolation threshold;

[0516] T_int = mean(I) ± 3 * std(I) / / reflection intensity threshold;

[0517] Where, I is the reflection intensity value;

[0518] R2. Threshold optimization:

[0519] Use ROC curve to analyze the performance of different thresholds;

[0520] Select a threshold value that balances the false positive rate and the false negative rate;

[0521] The threshold is spatially adaptively adjusted considering local terrain features;

[0522] After having the discrimination threshold, the system performs noise point recognition and removal operations:

[0523] J1. Calculate the comprehensive score for each point:

[0524] Score = w1* ISO + w2* |I mean(I)| / std(I);

[0525] Among them, w1 and w2 are weight coefficients, which are determined by cross-validation;

[0526] J2. Noise Marking:

[0527] If Score>T_combined (combined threshold), it is marked as noise;

[0528] Spatial clustering of points marked as noise points to avoid accidental deletion of small features;

[0529] J3. Noise Removal:

[0530] A progressive elimination strategy is adopted to eliminate noise points with high confidence first;

[0531] Save the removed point cloud information to support subsequent backtracking operations;

[0532] Integrity assessment is a key step to ensure the quality of noise processing:

[0533] Density analysis:

[0534] Calculate the change in point cloud density before and after removing noise points;

[0535] Detect whether there are areas with abnormal density;

[0536] Structural Integrity Check:

[0537] Analyze the continuity of key structural lines; evaluate the degree of preservation of feature boundaries; calculate local curvature changes;

[0538] Statistical indicators evaluation:

[0539] Calculate the skewness and kurtosis of point cloud distribution; analyze local flatness and roughness changes;

[0540] Finally, the system dynamically adjusts the discrimination threshold according to the integrity assessment results:

[0541] If a significant decrease in point cloud density is detected in certain areas, the system will:

[0542] Lower the isolation threshold of the area; relax the reflection intensity judgment conditions; increase the weight of local features;

[0543] If important structural features are found to be affected, the system will:

[0544] Improve the noise point discrimination threshold near the structure line; strengthen the constraints of local connectivity; introduce additional geometric features to assist judgment;

[0545] The threshold adjustment adopts an iterative optimization strategy:

[0546] Re-evaluate the treatment effect after each adjustment; record the optimal threshold combination; establish an experience database for threshold adjustment;

[0547] Through this adaptive noise processing mechanism, the system can effectively remove abnormal points in non-ground point clouds while maintaining the integrity of ground features, thereby improving the accuracy and reliability of subsequent modeling.

[0548] In one embodiment, the severe weather noise points are identified by using an improved MIS (maximum independent set) algorithm, and the specific process is as follows:

[0549] G1. Use the improved MIS algorithm to remove noise

[0550] Non-ground point clouds may contain noise introduced by bad weather, such as point cloud noise caused by raindrops, snowflakes or fog reflections. The system uses an improved MIS (maximum independent set) algorithm to identify these noise points.

[0551] The MIS algorithm makes judgments based on the spatial independence of the point cloud and can detect points with discontinuous distribution and strong isolation, which are often misdetected due to bad weather. For example, the point cloud reflected by raindrops is usually randomly distributed, highly independent, and incoherent with the surrounding terrain features. The system will identify and remove these noise points.

[0552] G2. Optimize the results of noise removal

[0553] The improved MIS algorithm can further remove false reflection point clouds in bad weather through iterative optimization. For example, the system can distinguish noise points from real non-ground point clouds based on their reflection intensity, spatial isolation and other characteristics, and remove them from the data set.

[0554] After each noise removal, the system will re-evaluate the overall structure of the point cloud data to ensure that the removed noise does not affect other real terrain features. After noise removal and data processing, the system will further analyze the remaining point cloud to identify key geographic information elements in karst landforms. The system will identify important features such as karst depressions, cave entrances, and steep cliffs through ground point clouds. Non-ground point clouds are used to identify related information such as trees and buildings. The system can track changes in karst landforms through time series data. For example, structural changes in caves, erosion of rock formations, etc., can be analyzed by comparing multiple surveying and mapping data. Specific severe weather conditions (such as heavy rains and floods) may have an impact on karst landforms. The system can identify terrain changes under these conditions and provide early warning services for related application scenarios.

[0555] S5: Using a hierarchical structure processing module to perform hierarchical optimization of ground point cloud and non-ground point cloud to generate three-dimensional terrain data;

[0556] In karst landform mapping, obstacles (such as complex rocks, caves, etc.) and bad weather (such as rain, fog, etc.) often bring noise to the lidar point cloud data, affecting the accuracy of the mapping data. In order to effectively deal with this noise problem, the use of multi-level structure and node weighting technology can provide an advanced connection marking scheme to separate and mark the noise introduced by obstacles and bad weather. This scheme is particularly suitable for the complex environment of karst landforms and can improve the accuracy of data analysis.

[0557] S5.1: Construct a hierarchical structure based on the spatial distribution characteristics of ground point clouds and non-ground point clouds;

[0558] S5.2: Divide the obstacle point cloud data into the first level;

[0559] S5.3: Classify the third noise data introduced by weather factors into the second level;

[0560] S5.4: Assign weights to nodes at each level according to the reflection intensity and density distribution of the point cloud;

[0561] S5.5: Optimize point cloud data based on node weights;

[0562] S5.5.1: Connect the markers of the continuously distributed obstacle point clouds to construct terrain contours;

[0563] S5.5.2: Eliminate abnormal point cloud data based on node weight;

[0564] S5.5.3: Dynamically adjust node weight coefficients according to different terrain features;

[0565] S5.5.4: Generate three-dimensional models of surface features.

[0566] After completing the classification of point cloud data, it is necessary to perform hierarchical optimization processing on the ground point cloud and non-ground point cloud to generate high-quality 3D terrain data. The hierarchical structure processing module adopts a hierarchical data organization method to optimize the processing of different types of point cloud data and finally generate an accurate 3D terrain model.

[0567] When constructing a hierarchical structure, the system first analyzes the spatial distribution characteristics of ground point clouds and non-ground point clouds. For ground point clouds, the main focus is on their elevation change characteristics, slope distribution, and roughness attributes; for non-ground point clouds, the focus is on analyzing their spatial aggregation characteristics, height distribution, and density changes. Based on these characteristics, the system uses a bottom-up approach to construct a multi-level data structure, and each level contains specific terrain information.

[0568] like Figure 4 As shown in the figure, in the first level, the system prioritizes obstacle point cloud data. These data usually include non-topographic elements such as buildings, vegetation, and artificial structures. During the processing, the obstacle point cloud is first grouped and clustered, and each cluster represents an independent obstacle entity. The system uses a density clustering algorithm to automatically determine the clustering parameters by analyzing the spatial density distribution of the point cloud. For building obstacles, it is also necessary to extract their contours and main structural surfaces, which are crucial for subsequent 3D modeling.

[0569] The second level mainly deals with the third noise data introduced by weather factors. This type of noise usually appears as randomly distributed discrete points, characterized by low density and irregular spatial distribution. The system identifies areas that may be affected by weather by analyzing the reflection intensity and echo characteristics of the point cloud. For example, data collected under foggy conditions may show abnormal reflection intensity, while under rainy conditions there may be multiple echoes. The system uses corresponding filtering algorithms to process different types of weather influences.

[0570] After completing the preliminary division of each level, the system needs to assign appropriate weights to the nodes of each level. The weight assignment is based on two main factors: the reflection intensity and density distribution of the point cloud. The reflection intensity can reflect the material characteristics of the target object. Generally, the reflection intensity of the ground point cloud is relatively stable, while the reflection intensity of non-ground point clouds such as vegetation fluctuates significantly. The density distribution reflects the reliability of the data, and areas with higher density usually have higher weights.

[0571] The system assigns weights to nodes in each layer (i.e., each point in the point cloud) based on the characteristics of each point cloud (such as reflection intensity, distribution density of adjacent points, etc.). Nodes with higher weights are more likely to be real point clouds of obstacles, while nodes with lower weights are more likely to be noise points.

[0572] Weighting mechanism: In karst landforms, real obstacle point clouds (such as rock walls, caves, etc.) usually have high spatial connectivity and stability, and these points will receive higher weights. Conversely, noise points such as raindrops and fog have irregular point cloud distribution and low density, and thus have lower weights.

[0573] The optimization process based on node weights is an iterative process. First, the system connects and marks the continuously distributed obstacle point clouds to construct a complete terrain outline. This process needs to consider the spatial relationship and morphological characteristics between the point clouds to ensure the rationality of the connection. For areas with data discontinuities, the system will interpolate and supplement based on the distribution characteristics of the surrounding point clouds.

[0574] During the processing, the system dynamically adjusts the node weight coefficients. This adjustment is based on the local characteristics of the terrain. For example, in areas with drastic slope changes, the system will increase the weight of the ground point cloud to maintain the accuracy of the terrain features; while in areas with dense vegetation, the weight of the non-ground point cloud needs to be increased to accurately reflect the surface coverage.

[0575] Finally, the system converts the optimized point cloud data into a 3D model. This process includes steps such as terrain meshing, surface reconstruction, and texture mapping. When generating a 3D model, the system uses an adaptive meshing strategy, using finer meshes for areas with drastic terrain changes and coarser meshes for flat areas, which can ensure the accuracy of the model while controlling the amount of data.

[0576] In order to ensure the quality of the 3D model, the system has also established a complete quality assessment mechanism. The assessment content includes the model's geometric accuracy, surface smoothness, degree of detail retention, etc. If it is found that the model quality of certain areas does not meet the requirements, the system will automatically mark these areas and give optimization suggestions.

[0577] Through this hierarchical optimization processing method, the system can effectively process point cloud data under complex terrain conditions and generate accurate 3D terrain models. These models not only accurately reflect the geometric characteristics of the surface, but also retain important ground feature information, which can provide reliable data support for subsequent terrain analysis and planning.

[0578] Through the multi-level structure and node weighting mechanism, the system connects and marks the first point cloud and the second point cloud. This process can connect the real obstacle point clouds belonging to the same level, while the noise points are separated and marked.

[0579] The system first marks the obstacle point clouds that are connected together to form a continuous terrain model.

[0580] Then, using the weighted results of the nodes, the system removes the noise points from the model or marks them separately to ensure that subsequent data analysis and feature mining are not interfered with.

[0581] like Figure 5 As shown, the embodiment of the present application also provides a multi-platform laser radar measurement device for implementing the above method, including:

[0582] A laser radar measurement host, wherein the laser radar measurement host includes a laser radar sensor, a Beidou navigation system and an inertial measurement unit, which are used to collect point cloud data, position data and attitude data respectively;

[0583] A multi-mode adapter device, including a rotor or fixed-wing bracket for an unmanned aerial vehicle platform, an adjustable bracket for a vehicle-mounted platform, and a portable backpack device for walking measurement, wherein the multi-mode adapter device is fixedly connected to the laser radar measurement host in a time-sharing manner;

[0584] A data processing module, the data processing module includes a planning perception diffusion network unit, an offline value function memory unit, a sequence exploration analysis unit and a secondary cutting plane processing unit arranged in series, and is used to perform hierarchical processing on the point cloud data;

[0585] A hierarchical structure processing module (not shown in the figure), which is connected to the data processing module via a data bus and is used to perform hierarchical optimization on the point cloud data;

[0586] The spatiotemporal big data platform is connected to the data processing module and the hierarchical structure processing module through a data interface, is used to store and manage data collected by multiple platforms, and provides data analysis and visualization functions.

[0587] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the implementation method of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation method of the present application.

[0588] Although the embodiments of the present disclosure are disclosed as above, the present invention is not limited thereto. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A multi-platform laser radar measurement method, characterized in that: The method is applied to the measurement of a laser radar measurement host in karst landforms, wherein the laser radar measurement host includes a laser radar sensor, a Beidou navigation system and an inertial measurement unit, and the method includes: Collecting point cloud data obtained by the laser radar measurement host under the UAV platform, vehicle-mounted platform and walking measurement platform respectively, wherein the laser radar measurement host is installed on the rotor or fixed-wing bracket of the UAV platform, on the adjustable bracket of the vehicle-mounted platform and in a portable backpack device for walking measurement in a time-sharing manner; Based on the real-time position data provided by the Beidou navigation system and the attitude data provided by the inertial measurement unit, the point cloud data is calibrated in space coordinates; The calibrated point cloud data is input into the planning perception diffusion network for first noise removal, and the point cloud data is classified by using the secondary quadratic cutting plane method to divide the point cloud data into ground point cloud and non-ground point cloud, wherein the planning perception diffusion network adopts an encoder-decoder structure, and realizes the recognition and removal of noise in the point cloud data through a multi-layer perceptron and an attention mechanism; Using a hierarchical structure processing module to perform hierarchical optimization of the ground point cloud and the non-ground point cloud to generate three-dimensional terrain data; The point cloud data is classified and processed by using the secondary quadratic cutting plane method to divide the point cloud data into ground point cloud and non-ground point cloud, including: Constructing multiple cutting planes based on the height information of the point cloud data and the spatial relationship between adjacent points; Mark the point cloud that is closely connected to the ground and has a height below a first threshold as a ground point cloud; Marking the point cloud separated from the ground or with a height higher than the first threshold as a non-ground point cloud; The ground point cloud and the non-ground point cloud are stored separately; The calibrated point cloud data is input into a planning perceptual diffusion network to remove the first noise point, including: Constructing terrain data segments according to the point cloud data, and inputting the terrain data segments into the planning perception diffusion network; Predicting a first noise point in a local area of ​​the terrain data segment by using the planning-aware diffusion network; Comparing the predicted first noise point information with the landform planning to verify the consistency between the first noise point information and the actual terrain features; After the verification is completed, the first noise point in the calibrated point cloud data is removed.

2. The method according to claim 1, characterized in that Before removing the first noise point in the calibrated point cloud data, the method further includes: Building a value function library based on historical surveying and mapping data and landform models, wherein the value function library stores point cloud noise processing strategies under different terrain scenarios; According to the characteristics of the current point cloud data, the optimal point cloud noise processing strategy is retrieved and called from the value function library; The optimal point cloud noise processing strategy is used to process the first noise point in the current point cloud data.

3. The method according to claim 1, characterized in that Before performing spatial coordinate calibration on the point cloud data, the method further includes: Organizing the point cloud data collected at different time series and different spatial positions according to a time-space sequence; Analyze point cloud data of different time series to extract long-term change trends of landforms; Based on the point cloud data, the landform structure features are analyzed from different spatial perspectives to explore the three-dimensional morphological change rules of the landform structure features; Based on the long-term change trend and the three-dimensional morphological change law, the point cloud data is preprocessed.

4. The method according to claim 1, characterized in that: The method further comprises: Constructing a connection relationship graph of the non-ground point cloud; Decomposing the nodes of the connection relationship graph using a deterministic network decomposition algorithm; identifying different types of features in the non-ground point cloud based on height distribution, density, and spatial relationships; analyzing tree and building features in the non-ground point cloud using a preset rule set; The second noise point in the non-ground point cloud is processed using a maximum independent set algorithm.

5. The method according to claim 4, characterized in that The second noise point in the non-ground point cloud is processed using a maximum independent set algorithm, including: Calculate the spatial independence parameters of point clouds and identify point clouds with discontinuous distribution; Analyze the reflection intensity and isolation degree of the point cloud to determine the discrimination threshold of the second noise point; Identifying and removing the second noise point based on the discrimination threshold; Performing integrity assessment on the non-ground point cloud data after removing the second noise point; The discrimination threshold of the second noise point is dynamically adjusted according to the integrity evaluation result.

6. The method according to claim 1, characterized in that A hierarchical structure processing module is used to perform hierarchical optimization of the ground point cloud and the non-ground point cloud, including: Constructing a hierarchical structure according to the spatial distribution characteristics of the ground point cloud and the non-ground point cloud; Divide the obstacle point cloud data into the first level; The third noise data introduced by weather factors is classified into the second level; Assign weights to nodes at each level based on the reflection intensity and density distribution of the point cloud; Point cloud data is optimized based on node weights.

7. The method according to claim 6, characterized in that Optimize point cloud data based on node weights, including: Connect and mark the continuously distributed obstacle point clouds to construct terrain contours; Eliminate abnormal point cloud data based on node weight; Dynamically adjust node weight coefficients according to different terrain features; Generate a 3D model of the Earth's surface features.

8. A multi-platform laser radar measurement device, characterized in that: The method for implementing any one of claims 1 to 7 comprises: A laser radar measurement host, wherein the laser radar measurement host includes a laser radar sensor, a Beidou navigation system and an inertial measurement unit, which are used to collect point cloud data, position data and attitude data respectively; A multi-mode adapter device, including a rotor or fixed-wing bracket for an unmanned aerial vehicle platform, an adjustable bracket for a vehicle-mounted platform, and a portable backpack device for walking measurement, wherein the multi-mode adapter device is fixedly connected to the laser radar measurement host in a time-sharing manner; A data processing module, the data processing module includes a planning perception diffusion network unit, an offline value function memory unit, a sequence exploration analysis unit and a secondary cutting plane processing unit arranged in series, and is used to perform hierarchical processing on the point cloud data; A hierarchical structure processing module, which is connected to the data processing module via a data bus and is used to perform hierarchical optimization on the point cloud data; The spatiotemporal big data platform is connected to the data processing module and the hierarchical structure processing module through a data interface, is used to store and manage data collected by multiple platforms, and provides data analysis and visualization functions.

Images