A human body three-dimensional structure modeling method based on single-station line-scan laser radar

Abstract

The present application relates to the technical field of laser radar point cloud processing and human posture reconstruction, in particular to a human three-dimensional structure modeling method based on single-station line-scan laser radar, in which, according to the determined effective layering meeting the human characteristics in the height direction and the offset of the center of gravity in the horizontal direction, continuity is judged, the effective layering meeting the continuity is clustered to obtain a human candidate region, the point cloud in the height range corresponding to the human candidate region is intercepted to form a human layered point cloud set, and the boundary range of the human in the height direction is determined. The present application does not need multiple sensors to cooperate or a complex learning model, and is suitable for human three-dimensional perception and structure analysis in complex environments.

Description

Technical Field

[0001] This invention relates to the field of lidar point cloud processing and human posture reconstruction technology, specifically a method for modeling the three-dimensional structure of the human body based on a monostation linear scan lidar. Background Technology

[0002] With the increasing prevalence of line-scan LiDAR in industrial safety monitoring, rail transportation, smart warehousing, and human behavior recognition, human detection and pose estimation technologies based on LiDAR point clouds have become a research hotspot. Existing technologies generally utilize cameras and deep learning models to infer key human points from images, or use 3D LiDAR point clouds to generate human skeletons. However, for cost-sensitive, complex environments, high privacy requirements, and applications that need to operate in real-time at the edge, traditional technologies still have significant limitations.

[0003] In online scanning LiDAR applications, human point cloud data is typically acquired in a layered manner along the height direction. Each layer contains a limited number of points, and their spatial distribution is uneven, making it difficult to form a complete and continuous three-dimensional volumetric structure. Existing human detection and pose estimation methods are mostly based on approaches such as overall point cloud clustering, single-layer contour analysis, or time-series tracking. These methods rely on relatively complete or balanced point cloud distribution conditions. However, in online scanning LiDAR applications with sparse and discrete point clouds, the stability of human region recognition and structural modeling is significantly limited.

[0004] In human body structure modeling, traditional methods typically estimate the location of the human body center or key points directly based on point cloud surface features. This lacks effective constraints on the spatial continuity between different height layers, easily leading to structural jumps in the height direction and making it difficult to form a stable and continuous description of the human body's central structure. Furthermore, methods that rely on multi-frame temporal information for compensation have limited effectiveness under single-frame or real-time processing requirements, making them unsuitable for deployment on edge devices or in environments with low computing resources.

[0005] Therefore, in human perception applications for line-scan lidar, there is an urgent need for a method that can fully utilize the spatial characteristics of layered point clouds and establish structural continuity relationships in the height direction to achieve stable localization of human body regions and continuous modeling of the central structure, so as to improve the reliability and practicality of human body structure extraction without the need for high-density point clouds or complex models. Summary of the Invention

[0006] The purpose of this invention is to provide a method for modeling the three-dimensional structure of the human body based on a single-station linear scanning lidar, so as to solve the problems mentioned in the background art.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a method for modeling the three-dimensional structure of the human body based on a single-station linear scanning lidar, the method comprising:

[0008] S1. Acquire layered point cloud data collected by the line-scanning lidar, perform denoising processing based on the spatial distance statistics of neighboring points on each layer of point cloud, and unify the three-dimensional coordinate system with the lidar emission center as the origin to obtain the denoised and coordinate-aligned layered point cloud; the specific process of the denoising processing is as follows: for each point in each layer of point cloud, find a preset number of nearest neighbor points through KD tree, calculate the average distance from the point to the nearest neighbor point, and then calculate the mean and standard deviation of the average distance of all points in the layer, and remove noise points with an average distance exceeding 2 times the standard deviation; the rules of the unified three-dimensional coordinate system are: the Z-axis is the height direction and is consistent with the lidar layer scanning direction, the X-axis is the horizontal left and right direction, and the Y-axis is the front and back direction and points to the side of the human body away from the lidar.

[0009] S2. Extract the multidimensional features of each layer of point cloud obtained in S1, and filter the layered point clouds based on the obtained multidimensional features to determine the effective layers that meet the human body characteristics. The multidimensional features include geometric features, echo intensity features and point cloud density features.

[0010] S3. Based on the spacing of the effective layers that meet the human body characteristics in the height direction and the offset of the center of gravity in the horizontal direction, the continuity is judged. The effective layers that meet the continuity are clustered to obtain human body candidate regions. The point cloud corresponding to the height range of the human body candidate regions is extracted to form a human body layered point cloud set, and the boundary range of the human body in the height direction is determined.

[0011] S4. For each layer of point cloud in the obtained human body layered point cloud set, the corresponding human body left and right structure midline is fitted by iterative optimization based on the point cloud symmetry characteristics. For abnormal layers with symmetry error exceeding the threshold, the midline position is corrected by interpolation.

[0012] S5. Divide the layered point cloud into left and right regions based on the central axis of the left and right structure, and adaptively estimate the front-back geometric dimensions of the human body in the corresponding layer by combining the distribution characteristics of the left and right region point clouds, and then determine the three-dimensional center point of the human body in the corresponding layer.

[0013] S6. Perform continuity constraints and smooth fitting on the three-dimensional center points of the human body in each layer to generate a three-dimensional central axis that runs through the height direction of the human body.

[0014] S7. Based on the obtained three-dimensional central axis and the inter-layer feature changes of the layered point cloud, determine the location of key points of the human body, and construct a human skeleton model according to the topological relationship of the human body.

[0015] Furthermore, the geometric features in S2 include the outline span and centroid position of the layered point cloud in the horizontal direction (the outline span of the layered point cloud in the horizontal direction of the X-axis and Y-axis, and the centroid position of the point cloud in the XY plane).

[0016] The echo intensity feature includes the statistical mean of the echo intensity corresponding to all points in the layered point cloud, which is used to distinguish the intensity difference between the human point cloud and the background point cloud; the point cloud density feature is the number of point clouds per unit volume, which characterizes the spatial distribution uniformity of the layered point cloud. It is calculated as follows: the effective volume of the cuboid is formed by extending a preset distance outward from the boundary of the point cloud outline, and the point cloud density is obtained by dividing the number of denoised point clouds by the effective volume.

[0017] The threshold condition for determining the effective layering of human characteristics in S2 is as follows: the average echo intensity of the layered point cloud is in the range of 200~800, and the point cloud density is ≥5 points / cm². 3 A stratification that meets both conditions is considered a valid stratification.

[0018] Furthermore, in S3, the number of consecutive effective layers corresponding to the human candidate region is greater than or equal to a preset threshold. The specific conditions for the continuity judgment in S3 are: the height direction distance between adjacent effective layers is ≤ the lidar layer height, the horizontal offset of the center of gravity on the X-axis is ≤ 3cm, and the horizontal offset of the center of gravity on the Y-axis is ≤ 3cm; when there are 5 or more consecutive effective layers (preset threshold) that meet the above continuity conditions, the consecutive layer cluster is determined to be a human candidate region; all layered point clouds within the height range from the lowest to the highest layer of the human candidate region are extracted to form a human layered point cloud set.

[0019] Furthermore, in S4, when determining the central axis of the left and right structures of the human body, the symmetry error of the layered point cloud about the candidate central axis is calculated, and the position of the central axis is optimized in an iterative manner; when the symmetry error of a certain layer exceeds a preset error threshold, interpolation correction is performed based on the central axis position of the adjacent layers.

[0020] The specific process of fitting and correcting the midline of the left and right structures of the human body in S4 is as follows:

[0021] S41. Using the X coordinate of the centroid of the layered point cloud outline as a reference, establish the initial left and right structural centerline of the human body extending along the Y axis.

[0022] S42. Calculate the mirror image of each point in the layered point cloud about the initial central axis, solve for the minimum distance from each mirror image to the original layered point cloud, and statistically obtain the average symmetry error of the corresponding layer.

[0023] S43. Adjust the X coordinate of the initial centerline with a preset iteration step size, and repeat step S42 until the average symmetry error converges to the preset accuracy to obtain the optimal centerline of the left and right human body structure for the corresponding layer.

[0024] S44. If the average symmetry error of a certain layer exceeds the preset error threshold, it is determined to be a symmetry abnormal layer. Select the optimal central axis X coordinate and corresponding height of the three effective layers adjacent to the corresponding abnormal layer, construct a quadratic polynomial and substitute it into the height of the abnormal layer to solve for the corrected optimal central axis X coordinate.

[0025] S45. The global symmetry axis reference value is obtained by statistically analyzing the layered central axis of the human torso region, and the local symmetry axis reference value is obtained by statistically analyzing the layered central axis of the preset key parts of the human body. If the deviation between the local reference value and the global reference value exceeds the set range, the central axis position of the corresponding area is corrected by using the local reference value.

[0026] Furthermore, the estimation of the geometric dimensions of the human body in the front-back direction in S5 is based on the lateral distance distribution of the point clouds in the left and right regions to the central axis, and is weighted and fused in combination with the symmetry of the distribution of the left and right point clouds, so as to achieve adaptive thickness compensation in the case of partial occlusion or posture deflection.

[0027] The specific execution steps are as follows:

[0028] S51. Based on the central axis of the human body's left and right structures, the layered point cloud is divided into a left point cloud and a right point cloud. If the number of point clouds on both sides is not lower than the preset lower limit, the subsequent steps are continued; otherwise, the corresponding layer is marked as an invalid layer and skipped.

[0029] S52. Calculate the lateral distance from each point in the left and right point clouds to the central axis. Use the 75th percentile to extract representative lateral distance values ​​for the left and right sides to reduce the impact of noise and local bulges on distance statistics.

[0030] S53. Calculate the consistency coefficient of the number of point clouds on the left and right sides. Based on the obtained consistency coefficient, perform weighted fusion on the representative lateral distance values ​​on the left and right sides, and adaptively estimate the front-back geometric dimensions of the human body in the corresponding layer (referring to the front-back dimensions of the human body from the front edge point of the human body to the plane containing the left and right midlines of the human body at the same height layer along the laser radar scanning direction (front-back Y-axis), i.e., the thickness of the human body).

[0031] The formula for calculating the consistency coefficient is as follows:

[0032] ;

[0033] in, is the consistency coefficient corresponding to the k-th layer of point cloud in the human body layered point cloud set; This is the ratio of the number of points on the left side to the total number of points on both sides within the k-th layer of the human body layered point cloud set; when When the value is 1, it indicates that the left and right point clouds in the k-th layer of the human body layered point cloud set are completely balanced; when When the value is less than 1, it indicates that the distribution of left and right point clouds in the k-th layer of the human body layered point cloud set is uneven.

[0034] The formula for calculating the anterior-posterior geometric dimensions of the human body is as follows:

[0035] ;

[0036] in: This is the estimated thickness of the k-th layer of point cloud in the human body layered point cloud set. This represents the representative horizontal distance value on the left side of the k-th layer of the human body point cloud set. `max()` represents the representative horizontal distance value on the right side of the k-th layer of the human body point cloud set; `max()` is the function to find the maximum value.

[0037] When the human body is relatively symmetrical, the thickness is taken as the average of the left and right sides;

[0038] When lateral tilt occurs (such as human body rotation), the thickness is automatically tilted to the larger side to ensure that the center point estimation is not distorted;

[0039] Then, using the X coordinate of the optimal central axis as the X coordinate of the three-dimensional center point, and the fixed height of the layered point cloud as the Z coordinate of the three-dimensional center point, combined with the geometric dimensions of the human body in the front-back direction and the coordinates of the front point of the layered point cloud on the Y axis, the Y coordinate of the three-dimensional center point is determined, thus obtaining the three-dimensional center point of the human body in this layer.

[0040] Furthermore, in step S6, when performing smooth fitting of the three-dimensional center point of the human body, the moving least squares method is used for local weighted fitting, so that the generated three-dimensional center axis of the human body remains continuous and smooth in the height direction and conforms to the human posture constraints.

[0041] The specific steps for obtaining the three-dimensional central axis of the human body are as follows:

[0042] S61. Calculate the three-dimensional Euclidean distance between each three-dimensional center point and the center points of the two adjacent layers before and after it. If the three-dimensional Euclidean distance between a certain center point and the adjacent layers before and after it is greater than the preset distance threshold, it is determined to be an isolated anomaly point. Its X and Y coordinates are corrected by linear interpolation, while the Z coordinate remains unchanged.

[0043] S62. The moving least squares method is used to perform local weighted fitting on all the corrected three-dimensional center points. A parameterized three-dimensional human body center axis is constructed with the height direction as the parameter. During the fitting process, the points that are closer to the current center point are given higher weights.

[0044] S63. Apply posture consistency constraints to the fitted central axis: In the standing posture, the angle between the direction vector of the central axis and the Z-axis is less than or equal to a preset angle; in the bending posture, the spatial curvature of the central axis is less than or equal to a preset curvature threshold. If the curvature exceeds the threshold, the fitting parameters are optimized to correct the axis shape.

[0045] Furthermore, in S7, when locating key points of the human body, candidate layers of key points are selected based on the interlayer differences in geometric shape, center of gravity position and echo intensity of adjacent layered point clouds, and multi-feature weighted fusion is performed on the point clouds in the candidate layers to determine the location of the key points.

[0046] Furthermore, the interlayer difference includes at least the relative change in contour span, the offset of the center of gravity, and the degree of change in strength, which are used to quantify the variation characteristics of human body structure in the height direction.

[0047] Furthermore, in S7, when constructing the three-dimensional skeleton model, the positions of key points are corrected based on the obtained three-dimensional central axis to make them closer to the actual joint positions inside the human body, and the key points are connected according to the human body topology to form a three-dimensional skeleton model.

[0048] The specific process of correcting the key point position based on the obtained three-dimensional central axis is as follows: the initial human body key point is offset along the X-axis direction towards the central axis direction, and the offset amount is a preset proportional coefficient of the distance from the initial key point to the central axis X-axis. The Y coordinate of the corrected key point is taken as the Y coordinate of the corresponding layered three-dimensional center point, while the Z coordinate remains unchanged, so that the corrected key point position is closer to the actual joint position inside the human body.

[0049] The specific process of constructing a three-dimensional skeleton model is as follows: Based on the human anatomical topology, the corrected key points of the human body are connected sequentially to form the core skeleton, upper limb skeleton, and lower limb skeleton; the core skeleton is the connection between the key points of the neck, shoulder, and hip; the upper limb skeleton is the connection between the key points of the shoulder, elbow, and wrist; and the lower limb skeleton is the connection between the key points of the hip, knee, and ankle. During the skeleton connection process, human kinematic constraints are applied, and reasonable ranges of values ​​are set for the joint angles to ensure the spatial structural rationality of the generated three-dimensional human skeleton model.

[0050] Furthermore, the method achieves continuous modeling and posture reconstruction of human body structure solely based on hierarchical point cloud data acquired by a single-station linear scan lidar, without relying on prior human body models, training data, or multi-sensor collaboration.

[0051] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0052] (1) This invention utilizes multi-layered point cloud data acquired by lidar along the height direction. First, it performs noise reduction and coordinate alignment on each layer of point cloud, and extracts geometric features, echo intensity features and point cloud density features to achieve stable identification of human body regions.

[0053] (2) This invention does not require multi-sensor collaboration or complex learning models and is suitable for human three-dimensional perception and structural analysis in complex environments. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart illustrating a method for modeling the three-dimensional structure of the human body based on a single-station linear scanning lidar according to the present invention.

[0056] Figure 2 This is a schematic diagram of background point cloud segmentation and pedestrian point cloud denoising processing in a human body three-dimensional structure modeling method based on a single-station linear scan lidar according to the present invention.

[0057] Figure 3 This is a schematic diagram of the left and right centerline positioning in a human body three-dimensional structure modeling method based on a single-station linear scanning lidar according to the present invention;

[0058] Figure 4 This is a schematic diagram of the positioning of the central symmetric point in a human body three-dimensional structure modeling method based on a single-station linear scanning lidar according to the present invention.

[0059] Figure 5 This is a schematic diagram of the human body thickness direction in a human body three-dimensional structure modeling method based on a single-station linear scanning lidar according to the present invention.

[0060] Figure 6 This is a schematic diagram of key points of the human body frame in the human body three-dimensional structure modeling method based on monostation line scan lidar of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1In this embodiment, the RoboSense RS-LiDAR-M1 MEMS line-scan lidar is used as the data acquisition device. Its core parameters are set as follows: ranging range 0.5~20m, layer height resolution 5cm, echo intensity output range 0~4095, scanning frequency 10Hz, and support for real-time output of the three-dimensional coordinates (X,Y,Z) and intensity values ​​(I) of the layered point cloud. The lidar is fixed on a 1.5m high tripod, with the transmission center horizontally facing the acquisition area, and large areas of highly reflective objects such as metal and glass are removed from the background.

[0063] S1. Acquire layered point cloud data collected by the line-scanning lidar, perform denoising processing based on the spatial distance statistics of neighboring points on each layer of point cloud, and unify the three-dimensional coordinate system with the lidar emission center as the origin to obtain the denoised and coordinate-aligned layered point cloud; the specific process of the denoising processing is as follows: for each point in each layer of point cloud, find a preset number of nearest neighbor points through KD tree, calculate the average distance from the point to the nearest neighbor point, and then calculate the mean and standard deviation of the average distance of all points in the layer, and remove noise points with an average distance exceeding 2 times the standard deviation; the rules of the unified three-dimensional coordinate system are: the Z-axis is the height direction and is consistent with the lidar layer scanning direction, the X-axis is the horizontal left and right direction, and the Y-axis is the front and back direction and points to the side of the human body away from the lidar.

[0064] Specifically, such as Figure 2 As shown, 100 subjects with different body types and heights (1.5-1.9m) were selected. They wore common clothing such as cotton and synthetic fibers and made everyday postures such as standing, slightly turning to the side, and slowly raising their hands. Each posture was collected for 5 seconds, and 457 sets of layered point cloud data (each set contains about 30 layers of point cloud) were obtained and stored in PLY format. Subject information, posture type and collection environment were recorded simultaneously.

[0065] A preprocessing program was developed based on the PCL 1.12 library. After reading PLY format data, the point cloud was split into layers according to the Z-axis height (the fixed height of the k-th layer is denoted as...). For example, when k=1 When k=2 The number of point clouds in the k-th layer is denoted as . Let the k-th layer point cloud be denoted as That is, the set of point clouds formed by the k-th layer. The KdTreeFLANN class is called to construct a KD-tree, setting the number of nearest neighbors to 5. For each point, its 5 nearest neighbors are searched, and then... The i-th point in the corresponding point cloud set is denoted as ;

[0066] Calculate according to the formula The average distance to the 5 nearest neighbors is denoted as . ;

[0067] ;

[0068] in, express The corresponding X-axis coordinate value; express The X-axis coordinate of the j-th nearest neighbor; express The corresponding Y-axis coordinate value; express The Y-axis coordinate of the j-th nearest neighbor;

[0069] Then calculate the mean of the average distances of all points in the corresponding stratum (denoted as ). ) and standard deviation (denoted as Set the noise removal threshold (the threshold is...). Eliminate those that satisfy the condition. For the noisy points in the condition, retain the k-th layer point cloud after denoising, denoted as . And update the number of point clouds in the k-th layer after denoising, denoted as . .

[0070] A unified three-dimensional coordinate system is established with the radar transmission center as the origin: the Z-axis represents the height direction, the X-axis represents the horizontal direction, and the Y-axis points to the side of the human body away from the radar. The `transformPointCloud` function is called to combine all the coordinates. Map to this coordinate system to ensure coordinate consistency between layers.

[0071] S2. Extract the multidimensional features of each layer of point cloud obtained in S1, and filter the layered point clouds based on the obtained multidimensional features to determine the effective layers that meet the human body characteristics. The multidimensional features include geometric features, echo intensity features and point cloud density features.

[0072] Specifically, for the point cloud after denoising at the k-th layer Extract the following three types of core features:

[0073] (1) Geometric features: traversal Extract the maximum / minimum values ​​on the X-axis, by Calculate X-axis span , express The maximum value on the X-axis in the traversal results. express The minimum value on the X-axis is obtained from the traversal results; similarly, the span on the Y-axis is obtained. , , express The maximum value of the Y-axis in the traversal results. express The minimum value of the Y-axis in the traversal results; according to the formula , Calculate the centroid of the k-th layer , This represents the X-axis coordinate of the centroid of the k-th layer. This represents the Y-axis coordinate of the centroid of the k-th layer. The Z-axis coordinate value of the centroid of the k-th layer is calculated by accumulating it in a loop.

[0074] (2) Intensity characteristics: Extract the intensity values ​​of all points in each layer, and denot the intensity value of the i-th point in the k-th layer as... ,according to Calculate the mean intensity of the corresponding layer (as the intensity feature of the corresponding layer), store it, and set the subsequent filtering threshold according to the radar intensity output range.

[0075] (3) Density characteristics: Based on the outline span, the effective area of ​​the cuboid is formed by extending 0.5cm along both the X and Y axes. (Z-axis thickness matched to layer resolution) Calculate volume. This represents the volume of the effective region of the cuboid formed after the expansion of the k-th layer; then, according to the formula... Calculate the point cloud density of the k-th layer. (Unit: points / cm) 3 Human point cloud density is typically distributed between 6 and 12 points / cm². 3 The background is less than 3 dots / cm 3 .

[0076] S3. Based on the spacing of the effective layers that meet the human body characteristics in the height direction and the offset of the center of gravity in the horizontal direction, the continuity is judged. The effective layers that meet the continuity are clustered to obtain human body candidate regions. The point cloud corresponding to the height range of the human body candidate regions is extracted to form a human body layered point cloud set, and the boundary range of the human body in the height direction is determined.

[0077] Specifically, an experiment was conducted to calibrate the threshold based on point clouds of 100 different human poses and background scenes: the intensity mean threshold. (Mean background intensity of metal is typically greater than 1000), density threshold dot / cm 3 After traversing all denoised layers, those that simultaneously meet both thresholds are marked as "valid layers". Add to candidate set .

[0078] Will Sort by Z-axis height from low to high, and calculate the centroid distance between adjacent effective layers (layer m and layer m+1): ( This represents the fixed height of the (m+1)th layer. This represents the fixed height of the m-th layer. (This represents the centroid spacing between the m-th effective layer and the (m+1)-th effective layer along the Z-axis). ; This represents the X-axis coordinate value of the centroid corresponding to the (m+1)th effective layer. This represents the Y-axis coordinate of the centroid corresponding to the (m+1)th effective layer. This represents the X-axis coordinate of the centroid corresponding to the m-th effective layer. This represents the Y-axis coordinate of the centroid corresponding to the m-th effective layer. This represents the centroid offset between the m-th effective layer and the (m+1)-th effective layer in the X-axis direction. This represents the centroid offset in the Y-axis direction between the m-th effective layer and the (m+1)-th effective layer.

[0079] Set continuity conditions: ( The radar layer height is represented by a value of 5cm, and the number of consecutive effective layers t≥5 (when the number of layers is less than 5, the false detection rate is high; the number of layers must reach a certain scale to reflect the spatial form of a complete vertical object like the human body). Continuous layer clusters that meet the conditions are selected as "human body candidate regions," denoted as G. Discrete background layers with t<5 are removed. , This represents the m-th valid layer within the continuous hierarchical cluster corresponding to the human candidate region.

[0080] The topmost height of the human candidate region G is extracted and denoted as . , (Corresponding to the position of the human head) and the height of the lowest layer, denoted as , (Corresponding to the position of the human foot);

[0081] Cut off height range All the layered point clouds within the body form a "human body layered point cloud set", denoted as This ensures that subsequent processing focuses only on the human body area, reducing invalid calculations.

[0082] S4. For each layer of the obtained human body layered point cloud set, based on the point cloud symmetry features, iteratively optimize and fit the corresponding layer's left and right human body midline. For abnormal layers with symmetry errors exceeding the threshold, interpolation is used to correct the midline position, such as... Figure 3 , Figure 4 As shown;

[0083] by each floor centroid X coordinate Based on this, establish the initial left and right midlines (the baseline for the left and right symmetry of the human body, the equation of the straight line). (Extending along the Y-axis), serving as the symmetry optimization benchmark; specifically: using the X-coordinate of the centroid as the initial symmetry axis position, such that this axis is positioned at the layer height. A straight line extending in the front-back direction is formed. This initial axis serves as the benchmark for subsequent symmetry optimization. In this paper, the left and right centerlines of the human body refer to the geometric centerlines that symmetrically divide the human body point cloud in the left and right directions within the same height layer, and their positions are determined through point cloud symmetry optimization.

[0084] For the k-th layer point cloud Points in Calculate its mirror image points about the initial left and right midlines of the human body, denoted as . , Call KD-tree search To the original layer point cloud The minimum distance is denoted as The formula is: .

[0085] The average symmetry error of the k-th layer is calculated using the formula and denoted as... , Set the iteration step size ΔX = 0.1cm and adjust the initial centerline X coordinate. Repeat the calculation of mirror points and symmetry errors until the errors converge to the set accuracy requirement (the difference in error between adjacent iterations < 0.05 cm). At this point, the midline is the optimal left and right midline of this layer, denoted as . , The corresponding X coordinates are .

[0086] A symmetry error threshold of 0.3 cm is set. If the average symmetry error of a certain layer is greater than 0.3 cm, the layer is determined to be an abnormal layer (such as the lateral upper limb layer). The optimal left and right midline X-coordinates of the human body are selected from the three adjacent effective layers (i.e., the effective layers 1 before and 2 after the abnormal layer, or 2 before and 1 after the abnormal layer). (Assuming k is the abnormal layer number, the optimal left and right midline X-coordinates of the effective layers corresponding to the two layers before and one layer after the abnormal layer are denoted as follows:) , , );

[0087] Constructing a quadratic polynomial Substitute the height Z of the adjacent effective layer and the corresponding X coordinates of the left and right midlines of the human body to solve for the coefficients a, b, and c.

[0088] Substituting the height of the anomaly layer into the polynomial, we obtain the corrected optimal symmetry axis X coordinate. .

[0089] Global symmetry axis fitting: The positions of multiple layers of symmetry axes in the main trunk area of ​​the human body are statistically analyzed to obtain the overall symmetry axis reference value; Local human body left and right midline fitting: Local symmetry axis reference values ​​are calculated for the height range corresponding to key parts of the human body (such as the upper trunk and waist and hip areas); A deviation threshold of 0.5cm is set. If the local and global deviations exceed the threshold, the midline of the corresponding area is corrected using the local value to enhance the positioning accuracy of the symmetry axis of key areas.

[0090] S5. Divide the layered point cloud into left and right regions based on the central axis of the left and right structure, and adaptively estimate the front-back geometric dimensions of the human body in the corresponding layer by combining the distribution characteristics of the left and right region point clouds, and then determine the three-dimensional center point of the human body in the corresponding layer.

[0091] Specifically, based on the X coordinate of the optimal midline of the k-th layer. Segmentation according to the formula: Denote the left-side point cloud as... , The point cloud on the right is denoted as , Set a minimum of 20 points for the point cloud. If the number of points on a certain side is less than 20, mark that layer as invalid and skip further processing.

[0092] calculate The horizontal distance from the X-axis coordinate value to the midline is denoted as . ;like belong Points inside, then ,like belong Points inside, then The 75th percentile was used to extract the representative distances to the left and right sides. The representative distance to the left side extracted from the 75th percentile was denoted as... The right-hand representative distance extracted from the 75th percentile is denoted as... This reduces the impact of noise and localized protrusions.

[0093] Left-right consistency: Quantified by calculating the ratio of the number of points on the left and right sides of the point cloud. The ratio of the number of points on the left side to the total number of points on both sides in the k-th layer of the human body layered point cloud set is denoted as... , The formula for calculating the consistency coefficient is as follows:

[0094] ;

[0095] in, is the consistency coefficient corresponding to the k-th layer of point cloud in the human body layered point cloud set; This is the ratio of the number of points on the left side to the total number of points on both sides within the k-th layer of the human body layered point cloud set; when When the value is 1, it indicates that the left and right point clouds in the k-th layer of the human body layered point cloud set are completely balanced; when When the value is less than 1, it indicates that the distribution of left and right point clouds in the k-th layer of the human body layered point cloud set is uneven, and there may be partial occlusion or posture deviation.

[0096] The formula for calculating the anterior-posterior geometric dimensions of the human body is as follows:

[0097] ;

[0098] in: This is the estimated thickness of the k-th layer of point cloud in the human body layered point cloud set. This represents the representative horizontal distance value on the left side of the k-th layer of the human body point cloud set. `max()` represents the representative horizontal distance value on the right side of the k-th layer of the human body point cloud set; `max()` is the function to find the maximum value.

[0099] When the human body is relatively symmetrical, the thickness is taken as the average of the left and right sides;

[0100] When lateral tilt occurs (such as human body rotation), the thickness is automatically tilted to the larger side to ensure that the center point estimation is not distorted;

[0101] Next, using the X-coordinate of the optimal central axis as the X-coordinate of the 3D center point, and the fixed height of the layered point cloud as the Z-coordinate of the 3D center point, combined with the geometric dimensions of the human body in the anterior-posterior direction and the coordinates of the front point of the layered point cloud along the Y-axis, the Y-coordinate of the 3D center point is determined, thus obtaining the 3D center point of the human body in that layer; the 3D center point of the human body in the kth layer is denoted as... ,

[0102] X direction: ;

[0103] Y-direction: Let the leading endpoint of this point cloud layer in the Y-direction be... Then the center point in the forward and backward directions is ;

[0104] Z direction: ;like Figure 5 As shown.

[0105] S6. Perform continuity constraints and smooth fitting on the three-dimensional center points of the human body in each layer to generate a three-dimensional central axis that runs through the height direction of the human body.

[0106] The core of this step is to correct the abnormal center point, smooth the fitted axis, and apply attitude constraints. The preferred solution is as follows:

[0107] (1) Screening and correction of anomalies in the center point

[0108] Calculate each The three-dimensional distance to the center point of the two adjacent layers is denoted as , This represents the three-dimensional center point of the human body in the (k-1)th layer;

[0109] ,

[0110] Set the 3D distance threshold to 5cm (based on the layered height resolution of the LiDAR and the characteristics of the continuous structure of the human body); if Simultaneously satisfy and ,determination For isolated outliers, linear interpolation is used for correction: The Z-coordinate remains unchanged.

[0111] (2) Smooth fitting of the central axis

[0112] To suppress the impact of local noise on the axial morphology while maintaining the continuity of the overall human body structure, the moving least squares (MLS) method is used for each... Five adjacent center points along the height direction are selected as the fitting window for local weighted fitting to construct the weights. By minimizing the error function The parameterized axis is obtained by fitting. , where the parameter t corresponds to the height direction of the human body.

[0113] (3) Attitude consistency constraints

[0114] Standing posture: The angle between the axial direction vector and the Z-axis is ≤15°;

[0115] Bending posture: Axial spatial curvature ≤ 0.02cm -1 If the value exceeds the threshold, the fitting parameters are optimized and the shape is corrected.

[0116] S7. Based on the obtained 3D central axis and the inter-layer feature changes of the layered point cloud, determine the positions of key points on the human body, and construct a human skeleton model according to the topological relationship of the human body, such as... Figure 6 As shown;

[0117] Specifically, according to the formula Calculate the interlayer variability. This represents the inter-layer variability corresponding to the k-th stratum, with a threshold of 0.05. If the value is 4095, it is marked as a candidate layer, and 4095 represents the maximum value of the selected lidar intensity.

[0118] Extract candidate points: geometric candidate points (top 20% of curvature), intensity candidate points (top 20% of intensity gradient), and density candidate points (top 20% of mesh density).

[0119] Weighted fusion: Geometric weight 0.5, intensity 0.2, density 0.3, calculate the comprehensive score, select the highest score as the keypoint, and denote the keypoint of the k-th layer as... .

[0120] Key point correction: Let γ = 0.3, then... The corrected X-axis coordinate value is denoted as Adjustment, The corrected Y-axis coordinate value is taken as follows , The corrected Z-axis coordinate value is taken as follows Connect the core skeleton (neck-shoulder-hip), upper limbs (shoulder-elbow-wrist), and lower limbs (hip-knee-ankle) according to the human body topology, apply joint angle constraints (elbow 30°~150°, knee 0°~170°), generate a three-dimensional skeleton model and save it.

[0121] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0122] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for modeling the three-dimensional structure of the human body based on a single-station linear scan lidar, characterized in that, The method includes: S1. Acquire layered point cloud data collected by line-scan lidar, perform denoising processing based on the spatial distance statistics of neighboring points on each layered point cloud, and unify the three-dimensional coordinate system with the lidar transmission center as the origin to obtain the denoised and coordinate-aligned layered point cloud. S2. Extract the multidimensional features of each layer of point cloud obtained in S1, and filter the layered point clouds based on the obtained multidimensional features to determine the effective layers that meet the human body characteristics. The multidimensional features include geometric features, echo intensity features and point cloud density features. S3. Based on the spacing of the effective layers that meet the human body characteristics in the height direction and the offset of the center of gravity in the horizontal direction, the continuity is judged. The effective layers that meet the continuity are clustered to obtain human body candidate regions. The point cloud corresponding to the height range of the human body candidate regions is extracted to form a human body layered point cloud set, and the boundary range of the human body in the height direction is determined. S4. For each layer of point cloud in the obtained human body layered point cloud set, the corresponding human body left and right structure midline is fitted by iterative optimization based on the point cloud symmetry characteristics. For abnormal layers with symmetry error exceeding the threshold, the midline position is corrected by interpolation. S5. Divide the layered point cloud into left and right regions based on the central axis of the left and right structure, and adaptively estimate the front-back geometric dimensions of the human body in the corresponding layer by combining the distribution characteristics of the left and right region point clouds, and then determine the three-dimensional center point of the human body in the corresponding layer. S6. Perform continuity constraints and smooth fitting on the three-dimensional center points of the human body in each layer to generate a three-dimensional central axis that runs through the height direction of the human body. S7. Based on the obtained three-dimensional central axis and the inter-layer feature changes of the layered point cloud, determine the location of key points of the human body, and construct a human skeleton model according to the topological relationship of the human body.

2. The method for modeling the three-dimensional structure of the human body based on a single-station linear scan lidar according to claim 1, characterized in that, The geometric features in S2 include the horizontal contour span and centroid position of the layered point cloud; The echo intensity feature includes the statistical mean of the echo intensity corresponding to all points in the layered point cloud; the point cloud density feature is the number of point clouds per unit volume.

3. The method for modeling the three-dimensional human body structure based on a single-station linear scan lidar according to claim 1, characterized in that, The number of consecutive effective layers corresponding to the human body candidate region in S3 is greater than or equal to a preset threshold.

4. The method for modeling the three-dimensional human body structure based on a single-station linear scan lidar according to claim 1, characterized in that, In step S4, when determining the central axis of the left and right structures of the human body, the symmetry error of the layered point cloud about the candidate central axis is calculated, and the position of the central axis is optimized in an iterative manner. When the symmetry error of a certain layer exceeds a preset error threshold, interpolation correction is performed based on the central axis position of the adjacent layers.

5. The method for modeling the three-dimensional structure of the human body based on a single-station linear scanning lidar according to claim 1, characterized in that, In step S6, when performing smooth fitting of the three-dimensional center point of the human body, the moving least squares method is used for local weighted fitting, so that the generated three-dimensional center axis of the human body remains continuous and smooth in the height direction and conforms to the human posture constraints.

6. The method for modeling the three-dimensional structure of the human body based on a single-station linear scan lidar according to claim 1, characterized in that, In S7, when locating key points of the human body, candidate layers of key points are selected based on the interlayer differences in geometric shape, center of gravity position and echo intensity of adjacent layered point clouds, and multi-feature weighted fusion is performed on the point clouds in the candidate layers to determine the location of the key points.

7. A method for modeling the three-dimensional structure of the human body based on a single-station linear scanning lidar according to claim 6, characterized in that, The interlayer variability includes at least the relative change in contour span, the offset of the center of gravity, and the degree of change in strength, and is used to quantify the variation characteristics of human body structure in the height direction.

8. A method for modeling the three-dimensional structure of the human body based on a single-station linear scan lidar according to claim 1, characterized in that, In S7, when constructing the three-dimensional skeleton model, the positions of key points are corrected based on the obtained three-dimensional central axis to make them closer to the actual joint positions inside the human body, and the key points are connected according to the human body topology to form a three-dimensional skeleton model.

9. A method for modeling the three-dimensional structure of the human body based on a single-station linear scan lidar according to claim 1, characterized in that, The method achieves continuous modeling and posture reconstruction of human body structure solely based on hierarchical point cloud data acquired by a single-station linear scan lidar, without relying on prior human body models, training data, or multi-sensor collaboration.

Images