Airborne laser radar point cloud inverted tree detection method and device and electronic equipment

Abstract

The invention discloses an airborne laser radar point cloud inverted tree detection method and device and electronic equipment, and the method employs a deep learning semantic segmentation model to process a first point cloud, can precisely distinguish an inverted tree trunk from ground cover objects, rocks and other sundries, remarkably reduces the false detection rate, and guarantees the accuracy of target recognition. And then, a two-dimensional plane projection image is generated based on the second point cloud, so that the problem of discontinuity caused by bending and fracture of the three-dimensional point cloud can be avoided, structured input is provided for subsequent target detection, and the continuity of the detection process is ensured. Afterwards, target detection is carried out on the two-dimensional plane projection image, fragmented missing detection caused by point cloud discontinuity can be avoided, and the overall recognition capability (high recall rate) of the inverted tree is greatly improved while high detection accuracy (low false positive) is kept. And finally, mapping a two-dimensional detection result back to a three-dimensional space, so that three-dimensional point cloud data of a complete inverted tree can be obtained, and a high-density point cloud environment can be adapted more effectively.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fallen tree detection, and particularly relates to an airborne laser radar point cloud fallen tree detection method and device and electronic equipment. BACKGROUND

[0002] In a forest ecosystem, fallen trees are an important indicator of biodiversity, forest health status, and the impact of natural disasters. Their spatial distribution information can not only be used to analyze forest succession patterns, but also provide key data support for post-disaster forestry resource management, fallen tree resource utilization, and ecological restoration monitoring.

[0003] In recent years, airborne laser radar (ALS) technology has become an important means for fallen tree detection in complex forest environments due to its ability to actively obtain high-precision three-dimensional point clouds. ALS point clouds can penetrate the canopy layer to directly obtain the structure of the fallen trees and the terrain under the trees, significantly improving detection efficiency and coverage.

[0004] However, the current detection method based on ALS point clouds still has the problems of low accuracy and insufficient completeness of the detection results. SUMMARY

[0005] In view of the above problems, the present application provides an airborne laser radar point cloud fallen tree detection method, device and electronic equipment to achieve the purpose of maintaining high accuracy (low false positive) while significantly improving the overall recognition ability of fallen trees (high recall rate). The specific scheme is as follows:

[0006] The first aspect of the present application provides an airborne laser radar point cloud fallen tree detection method, comprising:

[0007] obtaining a first point cloud, the first point cloud containing points in the original point cloud that satisfy at least a relative ground height within a preset threshold range;

[0008] processing the first point cloud based on a deep learning semantic segmentation model to obtain a second point cloud; the second point cloud contains points with a semantic of fallen tree trunks;

[0009] generating a two-dimensional plane projection image based on the second point cloud;

[0010] performing target detection on the two-dimensional plane projection image to obtain at least one fallen tree object;

[0011] extracting at least one target three-dimensional point set from the original point cloud based on the at least one fallen tree object.

[0012] In one possible implementation, the obtaining of the first point cloud comprises:

[0013] Collecting multiple point clouds in a preset forest area by using an airborne laser radar; the multiple point clouds correspond to different flight trajectories;

[0014] Based on the multiple point clouds, an original point cloud corresponding to the preset forest area is obtained;

[0015] The original point cloud is preprocessed to obtain a first point cloud.

[0016] In a possible implementation, the two-dimensional plane projection image includes a two-dimensional grid image; and the generating a two-dimensional plane projection image based on the second point cloud includes:

[0017] Performing a first translation operation on the second point cloud based on a set offset to obtain a translated second point cloud; the coordinate range of the translated second point cloud starts from 0 on each axis;

[0018] Extracting horizontal plane coordinates from the three-dimensional space coordinates of each point in the translated second point cloud;

[0019] Converting the horizontal plane coordinates into discrete pixel coordinates;

[0020] Creating a two-dimensional grid image according to the discrete pixel coordinates.

[0021] In a possible implementation, the airborne laser radar point cloud inverted tree detection method further includes:

[0022] Performing interpolation and / or smoothing processing on the two-dimensional grid image.

[0023] In a possible implementation, the target detection on the two-dimensional plane projection image to obtain at least one inverted tree object includes:

[0024] Inputting the two-dimensional plane projection image into a target detection model, and processing the two-dimensional plane projection image by the target detection model to obtain at least one bounding box; the bounding box represents the spatial distribution of the inverted tree trunk in the two-dimensional plane projection image.

[0025] In a possible implementation, based on the at least one inverted tree object, at least one target three-dimensional point set is extracted from the original point cloud, including:

[0026] Converting the pixel coordinates of the boundary points of the bounding box into horizontal plane coordinates of the boundary points;

[0027] Performing a second translation operation on the horizontal plane coordinates of the boundary points based on a set offset to obtain real horizontal plane coordinates of the boundary points;

[0028] Extracting a three-dimensional point set within a range determined by the real horizontal plane coordinates of the boundary points from the original point cloud as a target three-dimensional point set.

[0029] In a possible implementation, the deep learning semantic segmentation model is trained in the following manner:

[0030] obtain a sample point cloud;

[0031] input the sample point cloud into a deep learning semantic segmentation model to obtain a predicted semantic of each point in the sample point cloud determined by the deep learning semantic segmentation model;

[0032] adjust parameters of the deep learning semantic segmentation model based on a loss value, the loss value representing a difference between the predicted semantic of the point and a real semantic.

[0033] In a possible implementation, the airborne LiDAR point cloud inverted tree detection method further includes:

[0034] visualize the target three-dimensional point set;

[0035] determine at least one of a geometric morphological parameter, a shape, and a size of the inverted tree object based on the target three-dimensional point set;

[0036] output the at least one of the geometric morphological parameter, the shape, and the size of the inverted tree object.

[0037] In another aspect of the present application, an airborne LiDAR point cloud inverted tree detection device is provided, comprising:

[0038] an obtaining module configured to obtain a first point cloud, the first point cloud containing points in an original point cloud that at least satisfy a preset threshold range of relative ground height;

[0039] a processing module configured to process the first point cloud based on a deep learning semantic segmentation model to obtain a second point cloud, the second point cloud containing points with a semantic of inverted tree trunk;

[0040] a generating module configured to generate a two-dimensional plane projection image based on the second point cloud;

[0041] a detection module configured to perform target detection on the two-dimensional plane projection image to obtain at least one inverted tree object;

[0042] an extracting module configured to extract at least one target three-dimensional point set from the original point cloud based on the at least one inverted tree object.

[0043] In a third aspect of the present application, an electronic device is provided, comprising a memory, at least one processor, and a computer program stored in the memory, the processor executing the computer program to implement the following method steps:

[0044] Obtain a first point cloud, which contains points in the original point cloud that satisfy at least the condition that the relative ground height is within a preset threshold range;

[0045] Based on a deep learning semantic segmentation model, the first point cloud is processed to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks;

[0046] Based on the second point cloud, a two-dimensional planar projection image is generated;

[0047] Target detection is performed on the two-dimensional planar projection image to obtain at least one inverted tree object;

[0048] Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud.

[0049] In this application, the raw data is first preliminarily screened by acquiring a first point cloud, which effectively removes irrelevant point clouds, significantly reduces subsequent computational load, and lowers noise interference, laying the foundation for efficient processing. Next, a deep learning semantic segmentation model is used to process the first point cloud, accurately distinguishing fallen tree trunks from ground cover, rocks, and other debris, significantly reducing the false detection rate and ensuring the accuracy of target recognition. Subsequently, a two-dimensional planar projection image is generated based on the second point cloud. This image retains the planar contour information of the fallen tree trunk, avoiding the discontinuity problems caused by bending and breakage in the three-dimensional point cloud, providing structured input for subsequent target detection, and ensuring the continuity of the detection process. Afterwards, target detection is performed on the two-dimensional planar projection image, avoiding fragmented missed detections caused by point cloud discontinuity. While maintaining high detection accuracy (low false positives), the overall recognition ability of fallen trees is significantly improved (high recall). Finally, mapping the two-dimensional detection results back to three-dimensional space yields complete three-dimensional point cloud data of the fallen tree, enabling the entire method to more effectively adapt to high-density point cloud environments and achieve efficient and accurate fallen tree detection. Attached Figure Description

[0050] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0051] Figure 1 A flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 1 of this application;

[0052] Figure 2 This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 2 of this application.

[0053] Figure 3This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 5 of this application;

[0054] Figure 4 This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 6 of this application;

[0055] Figure 5 A flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 8 of this application;

[0056] Figure 6 This is a schematic diagram of the structure of an airborne lidar point cloud tree fall detection device provided in this application. Detailed Implementation

[0057] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0058] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0059] The terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0060] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] Reference Figure 1 This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 1 of this application. Figure 1 As shown, the method may include, but is not limited to, the following steps:

[0062] Step S101: Obtain a first point cloud, which contains at least the points in the original point cloud whose relative ground height is within a preset threshold range.

[0063] In this embodiment, the original point cloud may include: a three-dimensional spatial dataset acquired by an airborne lidar device, where each point may contain the following core attributes:

[0064] Three-dimensional coordinates (X, Y, Z): record the absolute position of a point in space;

[0065] Reflection intensity: Reflects the energy return value after the laser pulse interacts with the target surface, and is related to the target material (such as tree trunk, leaves, ground);

[0066] Other optional attributes (such as scanning angle, number of echoes, etc., depending on the device configuration)

[0067] In a forest scene, a threshold range (e.g., 0.2 meters to 5 meters) relative to ground height can be preset based on forest type, tree species, height, and fallen tree morphology (i.e., preset threshold range). This range can satisfy the following conditions:

[0068] Lower limit: The typical height above ground debris (such as fallen branches and stones) to avoid mistakenly including non-tree trunk targets;

[0069] Upper limit: Covering the main height of the fallen tree trunk (most fallen tree trunks are no more than 5 meters long), while excluding interference from upright tree crowns.

[0070] By obtaining the first point cloud, the computational load of subsequent deep learning semantic segmentation models can be significantly reduced, while also reducing the risk of overfitting due to data redundancy.

[0071] Furthermore, by filtering within a preset threshold range, interference from canopy point clouds and ground debris can be effectively suppressed, laying the foundation for high-precision detection.

[0072] Step S102: Based on a deep learning semantic segmentation model, process the first point cloud to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks.

[0073] In this embodiment, the three-dimensional coordinates and reflection intensity of each point in the first point cloud can be input into the deep learning semantic segmentation model to obtain the semantic labels of each point determined by the deep learning semantic segmentation model. The semantic labels can indicate whether the semantics of the point is an inverted tree trunk or a non-inverted tree trunk (e.g., noise or debris).

[0074] In this embodiment, points with semantic labels of inverted tree trunks can be extracted from the first point cloud, and the points with semantic labels of inverted tree trunks can be combined to form a second point cloud.

[0075] In this embodiment, the structure of the deep learning semantic segmentation model can be configured as needed. For example, the deep learning semantic segmentation model may include, but is not limited to, PointNet++, PointTransformer, or RandLA-Net.

[0076] Step S103: Generate a two-dimensional planar projection image based on the second point cloud.

[0077] In this embodiment, a two-dimensional planar projection image can be generated by projecting the second point cloud (which contains only points that are semantically inverted tree trunks) onto the horizontal plane (XY plane) as the projection reference and ignoring the height (Z axis) information.

[0078] The point cloud distribution of fallen trees in three-dimensional space often exhibits high dispersion due to their physical shape or acquisition conditions. For example,

[0079] After the tree trunk breaks into multiple segments, the point cloud of each segment is independently distributed in three-dimensional space (e.g., segment A is concentrated at a height of 1-2 meters, and segment B is concentrated at 2-3 meters), forming a clear spatial fault between the two segments.

[0080] When a tree trunk bends, the point cloud disperses along the curved path. In some areas (such as the inside of the bend), the point cloud may appear sparse due to the limited field of view of the sensor, forming "holes".

[0081] Obstacles such as leaves and buildings may completely block part of the tree trunk, preventing the sensor from collecting point clouds in the blocked area;

[0082] During tilted scanning, the distance between the distant tree trunk and the sensor increases, and the point cloud density decreases with the square of the distance, which may cause local sparsity (e.g., if the distance is doubled, the number of point clouds is reduced to 1 / 4).

[0083] To address the dispersion issue of 3D point clouds, they can be converted into 2D images through projection. During projection, all point clouds within the same XY plane region (regardless of height) can be compressed into a single pixel, achieving the aggregation of scattered point clouds and preserving the outline continuity of the fallen tree trunk. For example, if broken tree trunk segments overlap or are adjacent in the XY plane, their projected image will still appear as continuous bright areas, masking the gaps between the breaks. Alternatively, 3D bends can be converted into in-plane curves or polygonal lines, preserving the macroscopic extension direction of the tree trunk (e.g., from lower left to upper right) while ignoring local height variations (e.g., undulations at the bend).

[0084] Step S104: Perform target detection on the two-dimensional planar projection image to obtain at least one inverted tree object.

[0085] In this embodiment, the target detection model can be trained to learn the texture, shape, and contextual features of the fallen tree trunk (e.g., comparison with surrounding debris).

[0086] A two-dimensional planar projection image can be input into a trained object detection model, which will then process the image to obtain at least one inverted tree object.

[0087] The fallen tree object can include the fallen tree trunk and the possible branch extension area.

[0088] Step S105: Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud.

[0089] In this embodiment, at least one inverted tree object can be reverse-mapped to the original point cloud, and the corresponding 3D point set in the original point cloud can be used as at least one target 3D point set.

[0090] The target 3D point set corresponds to the inverted tree object and can contain all matching points of the inverted tree object in the original point cloud.

[0091] In this embodiment, the raw data is first preliminarily screened by acquiring a first point cloud, which effectively removes irrelevant point clouds, significantly reduces subsequent computational load, and lowers noise interference, laying the foundation for efficient processing. Next, a deep learning semantic segmentation model is used to process the first point cloud, accurately distinguishing fallen tree trunks from ground cover, rocks, and other debris, significantly reducing the false detection rate and ensuring the accuracy of target recognition. Subsequently, a two-dimensional planar projection image is generated based on the second point cloud. This image retains the planar contour information of the fallen tree trunk, avoiding the discontinuity problems caused by bending and breakage in the three-dimensional point cloud, providing structured input for subsequent target detection, and ensuring the continuity of the detection process. Afterwards, target detection is performed on the two-dimensional planar projection image, avoiding fragmented missed detections caused by point cloud discontinuity. While maintaining high detection accuracy (low false positives), the overall recognition ability of fallen trees is significantly improved (high recall). Finally, mapping the two-dimensional detection results back to three-dimensional space yields complete three-dimensional point cloud data of the fallen tree, enabling the entire method to more effectively adapt to high-density point cloud environments and achieve efficient and accurate fallen tree detection.

[0092] As another optional embodiment of this application, refer to Figure 2 This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 2 of this application. This embodiment mainly describes one implementation of step S101 above. Figure 2 As shown, step S101 may include, but is not limited to:

[0093] Step S1011: Use airborne lidar to collect multiple point clouds of a preset forest area; the multiple point clouds correspond to different flight trajectories.

[0094] In this embodiment, an airborne lidar device can be used to perform multiple flight scans of a preset forest area to collect multiple point cloud datasets covering different perspectives.

[0095] Multi-track acquisition compensates for blind spots in a single scan (such as the back of a tree trunk or depressions in the ground), improving the integrity and density of the point cloud. For example, in complex terrain, tilted flight trajectories can capture information about fallen trees beneath upright canopies, while vertical trajectories optimize the identification of ground debris (such as rocks).

[0096] Step S1012: Based on the multiple point clouds, obtain the original point cloud corresponding to the preset forest area.

[0097] In this embodiment, ICP (Iterative Closest Point) or NDT (Normal Distributions Transform) algorithms can be used to eliminate coordinate deviations between different trajectories and ensure spatial consistency of point clouds.

[0098] Then, point clouds of overlapping areas can be merged to avoid redundant calculations while preserving details in high-density areas (such as multi-segment point clouds of broken tree trunks).

[0099] Step S1013: Preprocess the original point cloud to obtain the first point cloud.

[0100] In this embodiment, the RANSAC (Random Sample Consensus) or CSF (Cloth Simulation Filtering) algorithms can be used to separate ground points from non-ground points in the original point cloud.

[0101] In this embodiment, a DTM (Digital Terrain Model) can be generated based on ground points to record the ground elevation distribution in the forest area (i.e., the Z value changes with the XY coordinates), providing a benchmark for subsequent normalization processing.

[0102] DTM can be used to calculate the height difference (ΔZ = Z_point - Z_DTM) for each non-ground point, which is the actual height of the point cloud relative to the ground.

[0103] In this embodiment, the height difference can be compared to see if it is within a preset threshold range, and the first point cloud that may belong to the fallen tree trunk (e.g., points whose ΔZ is within the preset threshold range) can be selected.

[0104] The first point cloud may contain fallen tree trunks, some low shrubs, and residual noise (such as birds and floating objects), which need to be further purified through subsequent semantic segmentation.

[0105] In this embodiment, by complementing data from multiple perspectives, the problem of blind spots in a single scan can be solved, thereby improving the integrity of the inverted tree point cloud and the robustness of detection.

[0106] Furthermore, by normalizing elevation, the influence of terrain undulation on the height threshold can be eliminated, making the preset threshold more universal and adaptable to different forest environments.

[0107] By setting a preset threshold range based on forest ecological characteristics, a balance can be struck between data retention and noise suppression, providing high-quality input for subsequent processing.

[0108] As another optional embodiment of this application, the airborne lidar point cloud inverted tree detection method provided in Embodiment 3 of this application is mainly an implementation of the above-mentioned step S103, and may include, but is not limited to, the following steps:

[0109] Step S11: Based on the set offset, perform a first translation operation on the second point cloud to obtain the translated second point cloud; the coordinate range of the translated second point cloud starts from 0 on each axis.

[0110] In this embodiment, the second point cloud can be traversed to find the minimum value of all points on the X, Y, and Z axes, i.e., X min =min(X1, X2, ..., X) n ), Y min =min(Y1, Y2, ..., Y) n ), Z min =min(Z1, Z2, ..., Z n ).

[0111] X min Y min and Z min It can be combined to get offset = (X) min Y min Z min ), offset can represent the set offset.

[0112] The coordinate range of the translated second point cloud starts from 0 on each axis, and the relative positional relationship of all points can remain unchanged.

[0113] Step S12: Extract the horizontal plane coordinates from the three-dimensional spatial coordinates of each point in the translated second point cloud.

[0114] In this embodiment, (X, Y) can be extracted from (X, Y, Z) of each point in the translated second point cloud, and the height information can be discarded.

[0115] Step S13: Convert the horizontal plane coordinates into discrete pixel coordinates.

[0116] In this embodiment, the pixel resolution (e.g., 5 cm / pixel) can be set according to the point cloud density and detection accuracy requirements. Then, for each horizontal plane head, it is converted into discrete pixel coordinates using the following formula:

[0117] pix=(xy / res+0.5).astype(“i4”).

[0118] pix can represent discrete pixel coordinates, xy represents horizontal plane coordinates, res represents pixel resolution, and astype(“i4”) represents rounding to the nearest integer.

[0119] Step S14: Create a two-dimensional raster image based on the discrete pixel coordinates.

[0120] In this embodiment, the image width (which can be represented as w) and height (which can be represented as h) can be determined based on the maximum value among the discrete pixel coordinates.

[0121] In this embodiment, a three-channel blank image (RGB format) with a size of h×w can be created, and the initial value can be black.

[0122] In this embodiment, pixel coordinates belonging to the inverted tree trunk can be filtered from each discrete pixel coordinate, and the filtered discrete pixel coordinates are marked with a set color (e.g., red) in the image to obtain a two-dimensional raster image.

[0123] In this embodiment, in order to retain more three-dimensional information in a two-dimensional image, the height Z can be encoded into the image's channels or grayscale values. For example, the height Z can be mapped to a channel (such as the B channel) of an RGB image, or the height Z can be converted to grayscale values ​​and then filled into a single-channel image.

[0124] Two-dimensional raster images are one implementation of two-dimensional planar projection images.

[0125] In this embodiment, a first translation operation is performed on the second point cloud. Since the coordinate range of the second point cloud in three-dimensional space may be quite dispersed, with significant differences in numerical values, this can introduce numerical error risks in subsequent calculations. By performing the first translation operation, the coordinate range of the point cloud can be normalized, ensuring that the value of each coordinate axis starts from 0. This significantly reduces the range of coordinate values, effectively mitigating the risk of errors caused by excessively large or small values ​​in subsequent calculations, and providing a more stable data foundation for subsequent processing.

[0126] After completing the first translation operation, a second point cloud is obtained. Next, horizontal coordinates (i.e., X-axis and Y-axis coordinates) are extracted from the 3D spatial coordinates of each point in this second point cloud. These horizontal coordinates are continuous real-valued coordinates; to facilitate subsequent image processing and to standardize data scale, they are converted into discrete pixel coordinates. During the conversion process, by setting an appropriate pixel resolution, the coordinates can be uniformly scaled and quantized.

[0127] Subsequently, a two-dimensional raster image is created based on the discrete pixel coordinates. Because the range of discretized pixel coordinates is typically scaled and quantized appropriately, its range is much smaller than the continuous coordinate range of the original point cloud on the horizontal plane. This means that the generated two-dimensional raster image covers a relatively small number of pixels, and the image size is much smaller than the data volume of the original point cloud. This reduction in data volume not only saves storage space but also greatly improves data processing efficiency.

[0128] After the above series of processing steps, the resulting 2D raster image has significant advantages. In 3D space, the point cloud of an inverted tree often exhibits highly dispersed and discontinuous characteristics due to physical morphological issues such as breakage, bending, or occlusion, as well as limitations in acquisition conditions. For example, broken tree trunk segments separate on the Z-axis, and the point cloud inside the bend is sparse, forming "holes." However, these problems are effectively solved in 2D raster images. All point clouds corresponding to the same XY coordinate can be compressed to the same pixel. If broken tree trunk segments overlap or are adjacent in the XY plane, they will appear as continuous bright areas in the image, masking the break gaps. 3D bends are converted into curves or polygons in the plane, preserving the macroscopic extension direction of the tree trunk while ignoring local height changes. Therefore, the inverted tree trunk appears as a continuous outline in the 2D raster image, avoiding the discontinuity problems in 3D space and providing more reliable and easier-to-process input data for subsequent target detection tasks.

[0129] As another optional embodiment of this application, the airborne lidar point cloud inverted tree detection method provided in Embodiment 4 of this application is mainly an implementation of the above-mentioned step S103, and may include, but is not limited to, the following steps:

[0130] Step S21: Based on the set offset, perform a first translation operation on the second point cloud to obtain the translated second point cloud; the coordinate range of the translated second point cloud starts from 0 on each axis;

[0131] Step S22: Extract the horizontal plane coordinates from the three-dimensional spatial coordinates of each point in the translated second point cloud;

[0132] Step S23: Convert the horizontal plane coordinates into discrete pixel coordinates;

[0133] Step S24: Create a two-dimensional raster image based on the discrete pixel coordinates.

[0134] For detailed procedures of steps S21-S24, please refer to the relevant description of steps S11-S14 in Example 3, which will not be repeated here.

[0135] Step S25: Perform interpolation and / or smoothing processing on the two-dimensional raster image.

[0136] Interpolation aims to use the pixel information around known pixels to insert new pixels into an image through specific algorithms (such as bilinear interpolation, cubic spline interpolation, etc.) to increase the resolution of the image, make the details in the image clearer, fill in the image information loss that may be caused by discretization processing, and thus more accurately present the morphological features of the fallen tree on the two-dimensional plane.

[0137] Smoothing involves adjusting pixel values ​​in the image using filtering algorithms (such as Gaussian filtering, median filtering, etc.) to remove noise and jagged edges, reduce irregular fluctuations, and make the image smoother and more natural overall, thereby improving image quality and providing more reliable and accurate input data for subsequent tree fall detection based on the two-dimensional raster image.

[0138] As another optional embodiment of this application, refer to Figure 3 This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 5 of this application. This embodiment is mainly an implementation of step S104 above, and may include, but is not limited to, the following steps:

[0139] Step S1041: Input the two-dimensional planar projection image into the target detection model, and process it to obtain at least one bounding box; the bounding box represents the spatial distribution of the fallen tree trunk in the two-dimensional planar projection image.

[0140] In this embodiment, a large number of two-dimensional planar projection images containing fallen tree scenes can be collected as a training dataset. These images can cover different forest types, tree species, fallen tree morphologies (such as bent, multi-segmented breaks, etc.), and lighting conditions to ensure the model's generalization ability.

[0141] The collected images are manually annotated, using rectangular bounding boxes to precisely define the fallen tree trunks and possible branch extensions within the images. During annotation, the bounding boxes should closely fit the outline of the fallen tree to avoid including excessive irrelevant background information.

[0142] In this embodiment, the object detection model may include, but is not limited to, the object detection network YOLOv8. YOLOv8 has high detection speed and high accuracy, and can quickly and accurately identify target objects in images, making it suitable for processing large-scale point cloud projection image data.

[0143] The labeled training data is input into the YOLOv8 model for training. During training, the model continuously adjusts the network parameters to learn the texture, shape, and contextual features of fallen tree trunks. For example, the texture of a fallen tree trunk is usually quite rough, showing a clear difference from the texture of surrounding debris (such as fallen leaves and stones); its shape may exhibit different forms such as long strips or curves; in terms of contextual features, the fallen tree trunk also differs from surrounding debris in brightness and contrast. The model learns these feature patterns to accurately identify fallen tree objects in subsequent detection.

[0144] The generated 2D planar projection image is input into a pre-trained YOLOv8 system. YOLOv8 divides the input 2D planar projection image into multiple uniform grids. Each grid is responsible not only for detecting objects but also for predicting a certain number of bounding boxes and their corresponding class probabilities. For each predicted bounding box, the model outputs its coordinate information (usually represented as the center coordinates (x, y), width w, and height h) and the probability value that the bounding box contains an inverted tree object. In this way, the model can perform object detection simultaneously on the entire image, avoiding the complex region proposal generation and filtering process in traditional methods, and greatly improving detection efficiency.

[0145] Classification and Positioning: Precisely Determining the Target

[0146] After completing the bounding box prediction, the model needs to classify and accurately locate each predicted bounding box to determine whether it contains the inverted tree object and accurately frame the target.

[0147] The classification process is based on the feature patterns of inverted tree objects learned by the model during the learning phase. The model matches and compares the image region within each bounding box with the learned features of the inverted tree object. By calculating similarity or probability values, the degree of matching between the image within the bounding box and the features of the inverted tree object is determined. If the matching degree is higher than a preset threshold, the bounding box is considered to contain an inverted tree object, and a corresponding category label (such as an inverted tree trunk) is given; otherwise, it is determined to be a non-inverted tree object (such as background or clutter).

[0148] The localization process involves precisely adjusting the bounding box coordinates using a regression algorithm. The initially predicted bounding box may not accurately enclose the inverted tree object due to various reasons (such as limitations in mesh generation or the complexity of the target shape). The regression algorithm fine-tunes the center coordinates, width, and height of the bounding box based on the target location information learned by the model, making it more closely match the actual outline of the inverted tree object, thus achieving accurate localization.

[0149] After processing by the object detection model, at least one bounding box can be obtained, which can represent the spatial distribution of the fallen tree trunk in the two-dimensional planar projection image. The bounding box is usually determined by four coordinate values ​​(such as the coordinates of the upper left and lower right corners) to determine its position and size in the image. These bounding boxes accurately enclose the fallen tree trunk and possible branch extension areas, solving the problem of overall detection of bent and multi-segmented fallen trees. For example, for a bent fallen tree trunk, the model can identify its overall extension direction and generate a bounding box that includes the entire bent trunk; for a multi-segmented broken fallen tree trunk, if the segments are adjacent or overlap in the two-dimensional projection image, the model can also detect them as a whole and generate a bounding box that covers all relevant segments.

[0150] In this embodiment, the target detection model, trained on a large amount of labeled data, learns rich features of fallen tree trunks, thereby reducing the misclassification of non-fallen objects as fallen trees during detection and maintaining high detection accuracy (low false positives). Simultaneously, because the two-dimensional projection image retains the macroscopic contour information of the fallen tree trunk, the model can more comprehensively detect actual fallen tree objects, significantly improving the overall recognition ability of fallen trees (high recall).

[0151] Furthermore, target detection models such as YOLOv8 have high detection speeds and can quickly process large-scale two-dimensional planar projection image data. Combined with the preliminary screening and semantic segmentation of the original point cloud in the previous steps, the entire airborne LiDAR point cloud fallen tree detection method can more effectively adapt to high-density point cloud environments and achieve efficient and accurate fallen tree detection.

[0152] As another optional embodiment of this application, refer to Figure 4This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 6 of this application. This embodiment is mainly an implementation of step S105 in Embodiment 5 above, and may include, but is not limited to, the following steps:

[0153] Step S1051: Convert the pixel coordinates of the boundary points of the bounding box into the horizontal plane coordinates of the boundary points.

[0154] Since the bounding box is obtained from a two-dimensional planar projection image, its boundary point coordinates are pixel coordinates. However, it needs to be matched with the original point cloud, which uses three-dimensional spatial coordinates. Therefore, the pixel coordinates of the bounding box boundary points must first be converted back to horizontal plane coordinates (two-dimensional spatial coordinates, corresponding to the X and Y coordinates in the original point cloud) in order to perform subsequent mapping operations.

[0155] It is known that when creating a two-dimensional planar projected image, the pixel resolution res is set, and the horizontal plane coordinates are converted to discrete pixel coordinates through a specific transformation relationship pix = (xy / res + 0.5).astype("i4") (where pix represents discrete pixel coordinates and xy represents horizontal plane coordinates).

[0156] Based on the reverse derivation of the above conversion formula, the pixel coordinates of the boundary points of the bounding box are converted into the horizontal plane coordinates of the boundary points.

[0157] Step S1052: Based on the set offset, perform a second translation operation on the horizontal plane coordinates of the boundary point to obtain the true horizontal plane coordinates of the boundary point.

[0158] In step S11, when performing the first translation operation on the second point cloud, the offset value offset = (X) was set. min , Y min Z min The point cloud coordinate range is normalized so that the value of each coordinate axis starts from 0. In this embodiment, a second translation operation, which is the reverse of the first translation operation, can be performed to map the horizontal plane coordinates of the boundary points back to the space of the original point cloud, restoring their true horizontal plane coordinate positions.

[0159] Step S1053: Extract the set of three-dimensional points within the range determined by the true horizontal plane coordinates of the boundary point from the original point cloud, and use it as the target three-dimensional point set.

[0160] In this embodiment, a range on the original point cloud horizontal plane can be determined based on the actual horizontal coordinates of the boundary points. For example, a rectangular range can be determined by calculating the maximum and minimum values ​​of the actual horizontal coordinates of the boundary points.

[0161] Based on the range determined by the true horizontal plane coordinates of the boundary points, all points in the original point cloud can be traversed. For each point (X, Y, Z), it is determined whether its horizontal plane coordinates (X, Y) are within the aforementioned defined rectangular range. If they are within the range, the point is added to the target 3D point set. The final target 3D point set corresponds to all 3D points of the inverted tree object detected in the 2D planar projection image within the original point cloud. These points contain the 3D spatial information of the inverted tree, such as height and shape, and can be used for further analysis, such as height measurement and morphological analysis of the inverted tree.

[0162] In this embodiment, by performing coordinate transformation, translation, and point set extraction operations, the accurate acquisition of the 3D information of the corresponding inverted tree object from the bounding box information in the 2D planar projection image to the original point cloud is achieved, providing reliable data support for the subsequent analysis and processing of the inverted tree object, and has significant beneficial effects.

[0163] As another optional embodiment of this application, the airborne lidar point cloud inverted tree detection method provided in Embodiment 7 of this application is mainly an implementation of the above-mentioned deep learning semantic segmentation model. The deep learning semantic segmentation model can be, but is not limited to, being trained in the following ways:

[0164] Step S31: Obtain the sample point cloud.

[0165] In this embodiment, an aircraft equipped with a lidar device (such as a drone or helicopter) can scan the target area according to a predetermined flight path and parameters. The lidar emits laser pulses towards the ground and receives the reflected signals. By measuring the flight time of the laser, the distance between the target object and the lidar is calculated, thereby obtaining the three-dimensional spatial information of the target area and forming the original point cloud data.

[0166] To improve the quality and richness of sample point clouds, other relevant data sources can be integrated. For example, optical imagery data can be combined, utilizing the rich color and texture information in the optical images to assist in the classification and labeling of point clouds. Image registration techniques can be used to precisely align the optical images with the LiDAR point clouds, ensuring that each point in the point cloud corresponds to its corresponding location in the optical image, thereby obtaining more feature information about ground features.

[0167] Raw airborne lidar point cloud data may contain noisy points, which may be caused by atmospheric scattering, multipath reflection, and other factors. Filtering algorithms are used to denoise the point cloud; common filtering algorithms include statistical filtering and radius filtering. Statistical filtering calculates the statistical characteristics (such as mean and standard deviation) of the surrounding neighborhood points for each point, and removes points that deviate from these statistical characteristics as noise. Radial filtering sets a radius threshold and counts the number of points in the surrounding neighborhood of each point; if the number of points is less than the set value, the point is considered noise and removed.

[0168] To meet the needs of fallen tree detection, different semantic categories can be defined. For example, points in a point cloud can be categorized into fallen trees, normal trees, buildings, and ground. The characteristics and criteria for each category should be clearly defined to ensure accurate labeling.

[0169] In this embodiment, professional point cloud annotation tools can be used, allowing professionals to semantically annotate each point in the sample point cloud. Annotators assign corresponding category labels to each point based on defined semantic categories, combined with information such as the geometric shape and spatial location of the point cloud. While manual annotation is time-consuming and labor-intensive, it ensures the accuracy and reliability of the annotation.

[0170] Step S32: Input the sample point cloud into the deep learning semantic segmentation model to obtain the predicted semantics of each point in the sample point cloud determined by the deep learning semantic segmentation model.

[0171] In this embodiment, the deep learning semantic segmentation model may include, but is not limited to:

[0172] Point-based convolutional models include PointNet and PointNet++. PointNet directly extracts features from each point in the point cloud, learning local and global features of points through a multilayer perceptron (MLP) network. It then aggregates the features of all points to achieve semantic segmentation of the point cloud. PointNet++ improves upon PointNet by introducing a hierarchical feature learning structure. Through sampling and grouping operations, it better captures the local spatial relationships of the point cloud, improving segmentation accuracy.

[0173] Graph neural network-based models treat points in a point cloud as nodes in a graph, and the spatial relationships between points as edges, constructing a point cloud graph structure. Then, a graph neural network (GNN) is used to propagate and aggregate features from the graph structure, learning contextual information between points to achieve semantic segmentation of the point cloud. For example, Dynamic Graph CNN (DGCNN) dynamically constructs neighborhood graphs between points and updates neighborhood relationships based on changes in point features, enabling better adaptation to local structural changes in the point cloud.

[0174] Step S33: Adjust the parameters of the deep learning semantic segmentation model based on the loss value; the loss value represents the difference between the predicted semantics and the true semantics (e.g., the category label mentioned above) of the point.

[0175] Adjusting model parameters is an iterative process, typically requiring multiple training iterations until the model converges. In each iteration, sample point clouds are input into the model for forward propagation, the loss value is calculated, and then the model parameters are adjusted using backpropagation and optimization algorithms. As the number of iterations increases, the model's performance gradually improves, and the loss value gradually decreases.

[0176] Several metrics can be used to determine whether a model has converged, such as changes in the loss value and the model's accuracy on the validation set. When the loss value changes little over multiple iterations, or when the model's accuracy on the validation set no longer improves significantly, the model can be considered to have converged, and training can be stopped.

[0177] In this embodiment, the deep learning semantic segmentation model can continuously learn and optimize, improving its semantic segmentation capability for different ground features such as fallen trees in airborne lidar point clouds, and providing accurate and reliable semantic information for subsequent fallen tree detection tasks.

[0178] As another optional embodiment of this application, refer to Figure 5 This is a flowchart illustrating an airborne lidar point cloud tree fall detection method provided in Embodiment 8 of this application. Figure 5 As shown, the method may include, but is not limited to, the following steps:

[0179] Step S201: Obtain a first point cloud, which contains at least the points in the original point cloud whose relative ground height is within a preset threshold range.

[0180] Step S202: Based on a deep learning semantic segmentation model, process the first point cloud to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks.

[0181] Step S203: Generate a two-dimensional planar projection image based on the second point cloud.

[0182] Step S204: Perform target detection on the two-dimensional planar projection image to obtain at least one inverted tree object.

[0183] Step S205: Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud.

[0184] For a detailed description of steps S201-S205, please refer to the relevant description of steps S101-S105 in Example 1, which will not be repeated here.

[0185] Step S206: Visualize the target three-dimensional point set.

[0186] Before visualization, preprocessing operations can be performed on the target 3D point set. For example, noise points can be removed to improve the quality of the point cloud and the visualization effect. Filtering algorithms, such as statistical filtering or radius filtering, can be used to remove noise points that deviate from the normal point distribution. Additionally, the point cloud can be downsampled to reduce the number of points and improve the rendering speed of visualization, especially when dealing with large-scale point clouds. Downsampling methods can include random downsampling or voxel mesh downsampling.

[0187] In this embodiment, the target 3D point cloud data can be imported into a professional-grade point cloud data processing platform. Users can observe the 3D point cloud morphology of the fallen tree object from different angles by adjusting the viewpoint, zoom level, rotation, and other operations. For example, the point distribution of the fallen tree trunk, branches, and other parts can be clearly seen, and the measurement tools can be used to perform preliminary measurements of geometric features such as distance and angle in the point cloud.

[0188] Step S207: Based on the target three-dimensional point set, determine at least one of the geometric morphological parameters, shape, and size of the inverted tree object.

[0189] In this embodiment, geometric parameters may include, but are not limited to, the length of the inverted tree object, the diameter of its cross-section, etc.

[0190] For example, for the trunk of a fallen tree, starting from the endpoint where the trunk contacts the ground, the line extends along the trunk's centerline (obtained by fitting a straight line or curve to the point cloud) to the other end, until the trunk reaches its end. In point cloud processing, principal component analysis (PCA) is first performed on the trunk's point cloud to find its principal direction, which approximates the trunk's orientation. Then, the coordinates of the two endpoints of the trunk along the principal direction are determined, and the trunk length (i.e., one implementation of the length of the fallen tree object) is obtained by calculating the Euclidean distance between these two endpoints.

[0191] For each branch of the inverted tree, the starting and ending points of the branch are determined first. The starting point of the branch is usually connected to the trunk or a superior branch, and the ending point is the end point of the branch. Using a method similar to that used for measuring the trunk length, the branch length (i.e., one implementation of the length of the inverted tree object) is obtained by calculating the Euclidean distance between the starting and ending points of the branch. After measuring all branches, the distribution of branch lengths can be obtained.

[0192] In this embodiment, cross-sections can be selected at different heights of the main trunk to measure the diameter of the cross-sections of the main trunk.

[0193] Alternatively, cross-sections can be selected at different locations on the branch to perform circular or elliptical fitting and calculate the diameter of the branch's cross-section. For thinner branches, the interval between measurement positions can be adjusted appropriately according to the point cloud density and accuracy requirements.

[0194] In this embodiment, the shape of the inverted tree object may include an overall shape and a local shape.

[0195] For example, if the trunk of a fallen tree is relatively straight and the branches are relatively regularly distributed, it can be preliminarily judged that its overall shape is a relatively regular tree shape; if the trunk is obviously bent and the branches are distributed in a disorderly manner, the overall shape may be an irregular twisted tree shape.

[0196] In this embodiment, simple geometric models (such as combinations of cylinders and cones) can be used to fit the fallen tree object. The suitability of the model is evaluated by calculating the fitting error. For example, if multiple cylinder combinations are used to fit the trunk and branches of the fallen tree, and the fitting error is within an acceptable range, the overall shape of the fallen tree can be judged to be similar to a shape formed by connecting multiple cylinders based on the combination of cylinders.

[0197] For the trunk, in addition to analyzing the cross-sectional shape (circle or ellipse) mentioned above, the degree of curvature in the longitudinal direction can also be observed. The curvature distribution of the trunk is obtained by calculating the change in tangent direction between adjacent points on the trunk. Regions with greater curvature indicate a greater degree of trunk bending, while regions with less curvature indicate a relatively straight trunk.

[0198] In this embodiment, the maximum length, width, and height of the inverted tree object in three-dimensional space can be determined. The overall dimensions can be obtained by finding the farthest points of the target three-dimensional points concentrated along the three coordinate axes and calculating the distances between them.

[0199] For each branch of the inverted tree, the length, diameter, and other dimensional parameters of each branch can be measured separately.

[0200] Step S208: Output at least one of the geometric parameters, shape, and size of the inverted tree object.

[0201] Choose the content to output based on the actual application requirements. If you are mainly concerned with the size information of the fallen tree, you can choose to output the overall dimensions such as length, width, and height, as well as local dimensions such as branch length and diameter. If you need to classify or analyze the shape of the fallen tree, you can choose to output the shape information. If you need to fully understand the characteristics of the fallen tree, you can output the geometric morphological parameters, shape, and size at the same time.

[0202] In this embodiment, 3D visualization graphics serve as a universal language for individuals with diverse professional backgrounds, such as forestry experts, data analysts, and engineers, when discussing issues related to fallen trees. It overcomes barriers of expertise, enabling parties to communicate and discuss based on intuitive graphics, conveying information more accurately and improving communication efficiency.

[0203] For example, in forestry disaster assessment projects, field surveyors, data processors, and decision-makers can work together to explore fallen tree situations and develop more reasonable response strategies through visualization.

[0204] Furthermore, the output results can provide important information for decision-making in related fields. In forestry management, the size and shape of fallen trees can determine whether they need to be cleared, how to clear them, and the priority of clearing. In forest resource assessment, geometric parameters can be used to calculate tree volume, providing data support for the rational utilization and protection of forest resources.

[0205] For example, after a forest fire, by outputting the parameters of fallen trees, the impact of the fire on the forest structure can be assessed, and a reasonable forest restoration plan can be formulated.

[0206] The following section introduces the airborne lidar point cloud fallen tree detection device provided in this application. The airborne lidar point cloud fallen tree detection device described below can be referred to in correspondence with the airborne lidar point cloud fallen tree detection method described above.

[0207] Reference Figure 6 The airborne lidar point cloud tree fall detection device includes: an acquisition module 100, a processing module 200, a generation module 300, a detection module 400, and an extraction module 500.

[0208] The module 100 is used to obtain a first point cloud, which contains at least points in the original point cloud whose relative ground height is within a preset threshold range.

[0209] The processing module 200 is used to process the first point cloud based on a deep learning semantic segmentation model to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks.

[0210] The generation module 300 is used to generate a two-dimensional planar projection image based on the second point cloud.

[0211] The detection module 400 is used to perform target detection on the two-dimensional planar projection image to obtain at least one inverted tree object.

[0212] Extraction module 500 is used to extract at least one target 3D point set from the original point cloud based on the at least one inverted tree object.

[0213] Module 100 can be obtained and can be used for:

[0214] Multiple point clouds were collected from a pre-defined forest area using an airborne lidar; the multiple point clouds corresponded to different flight trajectories.

[0215] Based on the multiple point clouds, the original point cloud corresponding to the preset forest area is obtained;

[0216] The original point cloud is preprocessed to obtain the first point cloud.

[0217] Module 300 can be generated and can be used for:

[0218] Based on a set offset, a first translation operation is performed on the second point cloud to obtain a translated second point cloud; the coordinate range of the translated second point cloud starts from 0 on each axis;

[0219] Extract the horizontal plane coordinates from the three-dimensional spatial coordinates of each point in the translated second point cloud;

[0220] Convert the horizontal plane coordinates into discrete pixel coordinates;

[0221] A two-dimensional raster image is created based on the discrete pixel coordinates.

[0222] Module 300 can also be used for:

[0223] The two-dimensional raster image is interpolated and / or smoothed.

[0224] The detection module 400 can be used specifically for:

[0225] The two-dimensional planar projection image is input into the target detection model, which processes it to obtain at least one bounding box; the bounding box represents the spatial distribution of the fallen tree trunk in the two-dimensional planar projection image.

[0226] Extraction module 500 can be used specifically for:

[0227] Convert the pixel coordinates of the boundary points of the bounding box into the horizontal plane coordinates of the boundary points;

[0228] Based on a set offset, a second translation operation is performed on the horizontal plane coordinates of the boundary point to obtain the true horizontal plane coordinates of the boundary point.

[0229] Extract the set of three-dimensional points within the range determined by the true horizontal plane coordinates of the boundary point from the original point cloud, and use it as the target three-dimensional point set.

[0230] In this embodiment, the airborne lidar point cloud tree fall detection device may further include:

[0231] The training module is used for:

[0232] Obtain sample point cloud;

[0233] The sample point cloud is input into a deep learning semantic segmentation model to obtain the predicted semantics of each point in the sample point cloud determined by the deep learning semantic segmentation model.

[0234] The parameters of the deep learning semantic segmentation model are adjusted based on the loss value; the loss value represents the difference between the predicted semantics and the true semantics of the point.

[0235] In this embodiment, the airborne lidar point cloud tree fall detection device may further include:

[0236] The visualization module is used to visualize the target three-dimensional point set.

[0237] The determination module is used to determine at least one of the geometric morphological parameters, shape, and size of the inverted tree object based on the target three-dimensional point set.

[0238] The output module is used to output at least one of the geometric parameters, shape, and size of the inverted tree object.

[0239] In another embodiment of this application, an electronic device is provided, which may include:

[0240] A memory, at least one processor, and a computer program stored in the memory, wherein the processor executes the computer program to perform the following method steps:

[0241] Obtain a first point cloud, which contains points in the original point cloud that satisfy at least the condition that the relative ground height is within a preset threshold range;

[0242] Based on a deep learning semantic segmentation model, the first point cloud is processed to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks;

[0243] Based on the second point cloud, a two-dimensional planar projection image is generated;

[0244] Target detection is performed on the two-dimensional planar projection image to obtain at least one inverted tree object;

[0245] Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud.

[0246] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0247] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0248] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0249] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for detecting fallen trees in point clouds using airborne lidar, characterized in that, include: Obtain a first point cloud, which contains points in the original point cloud that satisfy at least the condition that the relative ground height is within a preset threshold range; Based on a deep learning semantic segmentation model, the first point cloud is processed to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks; Based on the second point cloud, a two-dimensional planar projection image is generated; Target detection is performed on the two-dimensional planar projection image to obtain at least one inverted tree object; Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud.

2. The airborne lidar point cloud tree fall detection method according to claim 1, characterized in that, Obtaining the first point cloud includes: Multiple point clouds were collected from a pre-defined forest area using an airborne lidar; the multiple point clouds corresponded to different flight trajectories. Based on the multiple point clouds, the original point cloud corresponding to the preset forest area is obtained; The original point cloud is preprocessed to obtain the first point cloud.

3. The airborne lidar point cloud fallen tree detection method according to claim 1, characterized in that, The two-dimensional planar projection image includes: a two-dimensional raster image; generating the two-dimensional planar projection image based on the second point cloud includes: Based on a set offset, a first translation operation is performed on the second point cloud to obtain a translated second point cloud; the coordinate range of the translated second point cloud starts from 0 on each axis; Extract the horizontal plane coordinates from the three-dimensional spatial coordinates of each point in the translated second point cloud; Convert the horizontal plane coordinates into discrete pixel coordinates; A two-dimensional raster image is created based on the discrete pixel coordinates.

4. The airborne lidar point cloud tree fall detection method according to claim 3, characterized in that, The airborne lidar point cloud-based tree-falling detection method also includes: The two-dimensional raster image is interpolated and / or smoothed.

5. The airborne lidar point cloud fallen tree detection method according to claim 1, characterized in that, The step of performing target detection on the two-dimensional planar projection image to obtain at least one inverted tree object includes: The two-dimensional planar projection image is input into the target detection model, which processes it to obtain at least one bounding box; the bounding box represents the spatial distribution of the fallen tree trunk in the two-dimensional planar projection image.

6. The airborne lidar point cloud tree fall detection method according to claim 5, characterized in that, Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud, including: Convert the pixel coordinates of the boundary points of the bounding box into the horizontal plane coordinates of the boundary points; Based on a set offset, a second translation operation is performed on the horizontal plane coordinates of the boundary point to obtain the true horizontal plane coordinates of the boundary point. Extract the set of three-dimensional points within the range determined by the true horizontal plane coordinates of the boundary point from the original point cloud, and use it as the target three-dimensional point set.

7. The airborne lidar point cloud tree fall detection method according to claim 1, characterized in that, The deep learning semantic segmentation model is trained in the following way: Obtain sample point cloud; The sample point cloud is input into a deep learning semantic segmentation model to obtain the predicted semantics of each point in the sample point cloud determined by the deep learning semantic segmentation model. The parameters of the deep learning semantic segmentation model are adjusted based on the loss value; the loss value represents the difference between the predicted semantics and the true semantics of the point.

8. The airborne lidar point cloud fallen tree detection method according to claim 1, characterized in that, The airborne lidar point cloud-based tree-falling detection method also includes: Visualize the target 3D point set; Based on the target 3D point set, determine at least one of the geometric morphological parameters, shape, and size of the inverted tree object; Output at least one of the geometric parameters, shape, and size of the inverted tree object.

9. An airborne lidar point cloud tree-falling detection device, characterized in that, include: The module is used to obtain a first point cloud, which contains points in the original point cloud that at least satisfy the condition that the relative ground height is within a preset threshold range. The processing module is used to process the first point cloud based on a deep learning semantic segmentation model to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks; A generation module is used to generate a two-dimensional planar projection image based on the second point cloud; The detection module is used to perform target detection on the two-dimensional planar projection image to obtain at least one inverted tree object; An extraction module is used to extract at least one target 3D point set from the original point cloud based on the at least one inverted tree object.

10. An electronic device, characterized in that, The system includes a memory, at least one processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the following method steps: Obtain a first point cloud, which contains points in the original point cloud that satisfy at least the condition that the relative ground height is within a preset threshold range; Based on a deep learning semantic segmentation model, the first point cloud is processed to obtain a second point cloud; the second point cloud contains points whose semantics are inverted tree trunks; Based on the second point cloud, a two-dimensional planar projection image is generated; Target detection is performed on the two-dimensional planar projection image to obtain at least one inverted tree object; Based on the at least one inverted tree object, extract at least one target 3D point set from the original point cloud.

Images