Depth camera-based human body back spinal curvature abnormity identification method

Abstract

The invention relates to the technical field of spinal curvature recognition, in particular to a human body back spinal curvature anomaly recognition method based on a depth camera, and the method comprises the steps: fixing a laser depth camera through preset installation parameters, collecting an empty field point cloud, fitting a ground plane through an RANSAC algorithm, and building a mapping relation between pixels and physical coordinates. Furthermore, the positions of the center points of the back and the head are identified through a YOLOv8-s model, a spine center initial line is automatically generated, and stable detection in a natural posture is achieved. A chest area, a chest waist area and a waist area are divided based on a center line, key area point clouds are obtained by combining symmetrical sampling and filtering, and the accuracy and robustness of feature extraction are guaranteed. Extracting feature vectors of a tangent angle difference, an average height difference and a highest point height difference through height interpolation calculation and tangent fitting taking a ground plane as a reference; and finally, inputting the feature vectors into a full-connection neural network to complete scoliosis angle prediction and bending grade judgment, so that the detection precision and applicability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of spinal curvature recognition technology, and more particularly to a method for recognizing abnormal spinal curvature in the human back based on a depth camera. Background Technology

[0002] Abnormal spinal curvature, especially scoliosis, is a common skeletal development problem in adolescents. In its early stages, it often manifests as mild spinal deviation or scapular asymmetry. If it is not identified and intervened in time, it may lead to permanent spinal deformities and even respiratory and cardiovascular complications.

[0003] Existing spinal detection methods mainly fall into two categories: imaging detection and optical surface measurement. Imaging detection, such as X-rays or MRI, offers high diagnostic accuracy, but suffers from high costs and reliance on medical facilities, making it unsuitable for large-scale screening among adolescents. While optical measurement methods enable non-contact detection, most systems rely on single RGB images or two-dimensional depth data, lacking precise spatial scale calibration and failing to reflect subtle changes in the three-dimensional morphology of the spine. Especially in complex environments such as natural lighting, clothing obstruction, and postural deviations, traditional algorithms are susceptible to noise, leading to unstable recognition results. Existing spinal detection methods generally rely on key point extraction from two-dimensional images, making it difficult to combine physical coordinates for accurate three-dimensional modeling, and failing to establish a mapping relationship between pixels and physical space, resulting in errors in height and angle estimation. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for identifying abnormal curvature of the human spine based on a depth camera.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying abnormal curvature of the human back spine based on a depth camera includes the following steps: S1. Fix the laser depth camera according to the preset installation parameters, collect point cloud data of the open field, perform voxel downsampling on the point cloud and use the RANSAC algorithm to fit the ground plane, calculate the vertical height of the camera to the ground according to the fitting equation, and obtain the mapping relationship between pixels and physical coordinates. S2. Based on the mapping relationship, synchronously acquire RGB images and point cloud data of the detection object, input the images into the YOLOv8-s model to identify the center points of the back and head, and calculate the line connecting the two points to obtain the initial line of the spine center. S3. Based on the initial line of the spine center, the back is divided into three segments: chest, thoracolumbar and lumbar. The left and right symmetrical sampling windows are determined according to the pixel physical mapping. The corresponding point cloud subsets are extracted and filtered to obtain point cloud data of key areas. S4. Based on the point cloud data of the key area, calculate the left and right height distribution with the ground plane as the reference, interpolate the height values ​​to generate a height curve, locate the left and right highest points in the curve and fit the tangent line to obtain the tangent angle difference, average height difference and highest point height difference. S5. Input the feature vectors of the tangent angle difference, average height difference and highest point height difference into the fully connected neural network model, output the scoliosis angle, and determine the degree of spinal curvature according to the preset range of the scoliosis angle.

[0006] Further, S1 includes the following steps: S11. Based on the camera pose calibration parameters, obtain the intrinsic and extrinsic parameter matrices of the laser depth camera, perform registration calculations on the camera optical axis direction and the ground normal vector, and obtain the unified transformation matrix between the camera coordinate system and the world coordinate system. S12. Based on the unified transformation matrix, voxel downsampling is performed on the airfield point cloud data, the centroid coordinates of the point cloud within the voxels are calculated and the mean point cloud set is output; RANSAC plane fitting is performed on the mean point cloud set to obtain the ground plane equation and the set of points in the plane. S13. Calculate the vertical height of the camera from the ground based on the ground plane equation and the coordinates of the camera's optical center, and use this height as a scale reference to normalize the point cloud coordinates to obtain spatially normalized point cloud data. S14. Based on the normalized point cloud data and the camera intrinsic parameter matrix, perform three-dimensional back projection and plane intersection calculation on the pixel coordinates to generate a corresponding sample set of pixel points and ground physical coordinates; fit a mapping function according to the sample set to obtain the mapping relationship between pixel coordinates and physical coordinates.

[0007] Further, performing RANSAC plane fitting on the mean point cloud to obtain the ground plane equation and the set of points in the plane includes the following steps: S131. Based on the set of mean points, three points located in different spatial positions and not collinear are randomly selected from them. They are substituted into the plane equation, and the plane parameters of the current round are obtained by solving the system of linear equations. S132. Based on the plane parameters, calculate the Euclidean distance from all points in the mean point cloud to the plane, and set the distance threshold to 0.01m. Define points whose distance does not exceed the threshold as interior points, and count the number of interior points in the current round. S133. Repeat steps S131-S132 500 times, and select the plane with the most interior points and an interior point ratio of not less than 95% as the optimal ground plane. S134. Retain the plane equation and its interior point set corresponding to the optimal ground plane.

[0008] Furthermore, the step of synchronously acquiring the RGB image and point cloud data of the detection object according to the mapping relationship includes the following steps: Based on the mapping relationship between pixels and physical coordinates, external parameter calibration and time synchronization are performed on the laser depth camera and the RGB camera to establish a unified timestamp and spatial coordinate system and obtain the registration parameters between the cameras. Based on the registration parameters, multiple frames of RGB image data and depth point cloud data of the detected object within the acquisition area are obtained, and the camera pose and exposure parameters are recorded simultaneously during the acquisition process.

[0009] Furthermore, the YOLOv8-s model is trained through the following steps: The pre-acquired image data is labeled with back and head, and then size normalization, brightness enhancement and rotation perturbation are performed to obtain the enhanced training sample set. Based on the training sample set, a YOLOv8-s detection network model is constructed. Feature extraction and target classification training are performed on the input image. Initial detection weight parameters are output through the cosine annealing learning rate scheduling mechanism and the Adam optimization algorithm. Transfer learning is performed based on the initial detection weight parameters. The backbone network is fixed and only the detection head is trained to obtain the optimized weights of the detection head. Then, the entire network is fine-tuned and trained to output the converged complete model weights. Based on the complete model weights, the model is converted into the OpenVINO intermediate representation format, and the parameters are quantized using FP16 to obtain a lightweight YOLOv8-s model.

[0010] Furthermore, the pre-collected image data includes human sample images of different genders, ages, heights, body types, clothing types, lighting conditions and background environments, and posture deviations.

[0011] Further, S3 includes the following steps: Based on the initial line of the spine center and the mapping relationship between pixels and physical coordinates, the back region of the detected object is spatially projected, and the vertical projection range of the spine center line in the physical coordinate system is divided into chest, thoracolumbar and lumbar regions to obtain the corresponding region boundary coordinates. Based on the boundary coordinates of the region, a left-right symmetrical sampling window is constructed in each segmented region. Three-dimensional back projection calculation is performed on the pixel coordinates in each sampling window to obtain the candidate point cloud datasets on the left and right sides. Based on the point cloud candidate dataset, statistical filtering and outlier removal are performed on the point cloud of each segment region. The point cloud density distribution and spatial variance are calculated, and effective point cloud data after noise suppression are obtained. Based on the effective point cloud data, according to the segmented regions of chest, chest-waist, and waist and their left-right symmetry, a point cloud subset with uniform local spatial features is extracted and combined to obtain the key region point cloud dataset.

[0012] Further, S4 includes the following steps: Based on the point cloud data of the key area and the ground plane equation, the vertical distance from the point to the plane is calculated for the left and right point clouds of each segment area to obtain the corresponding height distribution dataset; Based on the height distribution dataset, the height values ​​of the left and right point clouds in each region are sorted and linearly interpolated according to the horizontal index to generate continuous left and right height curves. Based on the left and right height curves, locate the highest left and right points in each region, and extract local point cloud data in the neighborhood of the highest point; Based on the local point cloud data, the least squares method is used to fit the tangents within ±5mm of the highest point to obtain the left and right tangent angle values. Based on the left and right tangent angle values ​​and the left and right height curves, calculate the tangent angle difference, average height difference, and highest point height difference for each region.

[0013] Further, S5 includes the following steps: Based on the data of tangent angle difference, average height difference and highest point height difference, multidimensional feature vectors are constructed according to the chest, chest-waist and waist regions. Each feature is input into a trained fully connected neural network model, and multi-layer linear mapping and non-linear activation calculations are performed on the input data to obtain the predicted values ​​of spinal curvature in each region. The global scoliosis angle is calculated based on the predicted spinal curvature value, and the prediction results of different regions are fused by weighted averaging to output the overall spinal scoliosis angle. The spinal curvature level is determined and output by matching the overall scoliosis angle with a preset level threshold table.

[0014] Furthermore, the fully connected neural network is constructed through the following steps: A training sample set is established based on the multidimensional feature vector and the corresponding actual Cobb angle data of the spine, and the input features are normalized to obtain standardized feature data. Based on the standardized feature data, the number of neurons in the input layer is set to be consistent with the feature dimension, and a fully connected network structure containing two hidden layers is constructed. Each hidden layer is activated by the ReLU function and Dropout is set to suppress overfitting to obtain the initial network model. Based on the initial network model and training sample set, iterative training is performed using the mean squared error loss function and the Adam optimization algorithm. Gradient updates are performed on the network weight parameters until the loss converges, and the converged model parameters are obtained. Based on the converged model parameters, a prediction calculation is performed on the validation data, the predicted scoliosis angle is output, and the network parameters that meet the accuracy threshold are solidified into a fully connected neural network that has been trained.

[0015] The beneficial effects of this invention are as follows: By fixing a laser depth camera and acquiring point cloud data in an empty field, this invention uses the RANSAC algorithm to fit the point cloud to the ground plane, establishing a precise mapping relationship between pixel coordinates and physical coordinates. This fundamentally solves the problem of spatial scale calibration that cannot be achieved in traditional optical detection, providing a unified physical measurement basis for subsequent 3D morphological analysis. Furthermore, by introducing an optimized YOLOv8-s depth detection model, it can stably identify the center point positions of the back and head under complex lighting and diverse clothing backgrounds, automatically generating an initial line for the spine center, significantly improving the robustness of pose detection in natural scenes. Based on the initial line for the spine center, the back region is divided into three segments: chest, thoracolumbar, and lumbar. A symmetrical sampling window is constructed in the 3D point cloud space, and key area point cloud data is obtained through filtering and outlier removal, ensuring the stability and accuracy of the feature extraction process under different body shapes and poses. In the feature extraction stage, the point cloud height distribution is interpolated using the ground plane as a reference to generate left and right height curves. Geometric feature parameters such as the difference in tangent angles, average height difference, and highest point height difference are extracted. This effectively overcomes the limitation of traditional two-dimensional image analysis in reflecting the depth changes of the back curvature, achieving three-dimensional structural quantification of spinal curvature features. Finally, the extracted feature vectors are input into a trained fully connected neural network model. Through multi-layer mapping and nonlinear activation, the scoliosis angle is automatically predicted, and the curvature level is determined based on the angle range. This achieves intelligent recognition from three-dimensional shape to curvature degree, significantly improving detection accuracy and judgment consistency. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps in a method for identifying abnormal curvature of the human back spine based on a depth camera, as described in this invention.

[0017] Figure 2 This is a flowchart of step S3 in this invention. Detailed Implementation

[0018] Please see Figures 1-2 As shown, this invention relates to a method for identifying abnormal curvature of the human spine based on a depth camera, comprising the following steps: S1. Fix the laser depth camera according to the preset installation parameters, collect point cloud data of the open field, perform voxel downsampling on the point cloud and use the RANSAC algorithm to fit the ground plane, calculate the vertical height of the camera to the ground according to the fitting equation, and obtain the mapping relationship between pixels and physical coordinates. S2. Based on the mapping relationship, synchronously acquire RGB images and point cloud data of the detection object, input the images into the YOLOv8-s model to identify the center points of the back and head, and calculate the line connecting the two points to obtain the initial line of the spine center. S3. Based on the initial line of the spine center, the back is divided into three segments: chest, thoracolumbar and lumbar. The left and right symmetrical sampling windows are determined according to the pixel physical mapping. The corresponding point cloud subsets are extracted and filtered to obtain point cloud data of key areas. S4. Based on the point cloud data of the key area, calculate the left and right height distribution with the ground plane as the reference, interpolate the height values ​​to generate a height curve, locate the left and right highest points in the curve and fit the tangent line to obtain the tangent angle difference, average height difference and highest point height difference. S5. Input the feature vectors of the tangent angle difference, average height difference and highest point height difference into the fully connected neural network model, output the scoliosis angle, and determine the degree of spinal curvature according to the preset range of the scoliosis angle.

[0019] In some embodiments, the laser depth camera is fixedly mounted using an adjustable bracket. The vertical distance between the camera's optical center and the ground is 2.2m, the lens axis is perpendicular to the ground, and the detection area is a 1.0m × 0.8m rectangle. First, empty field point cloud data (resolution 1280 × 720, frame rate 30fps, point cloud density ≥ 300 points / cm²) is acquired, and voxel downsampling is performed on the point cloud to reduce noise density. Ground plane fitting is performed using the RANSAC algorithm, setting a distance threshold of 0.01m and iterating 500 times to select a plane with an inlier ratio ≥ 95% as the optimal ground plane. The vertical height from the camera to the ground is calculated based on the fitted plane equation and the camera's optical center coordinates, thus establishing a mapping function between pixel coordinates and physical coordinates, providing a unified spatial calibration benchmark for subsequent 3D morphological analysis. This step differs from existing 2D image methods, eliminating depth offset errors and achieving physical quantification of the measurement. Subsequently, RGB images and point cloud data of the detection object are simultaneously acquired according to the mapping relationship. The images are input into a YOLOv8-s model optimized through transfer learning to identify the center points of the back and head, generating an initial line for the spine center. Compared to traditional convolutional detection structures, this model adds a C2f module to the feature extraction layer to improve target boundary localization accuracy. It is then deployed to embedded systems using FP16 quantization for real-time detection and highly robust recognition. In the point cloud analysis stage, the back region is divided into three segments—chest, thoracolumbar, and lumbar—based on the initial centerline. A symmetrical sampling window is determined by combining pixel-to-physical coordinate mapping. Key regions are extracted from the point cloud, and statistical filtering is performed to remove outliers and random noise. The filtered key point cloud is used to calculate its height distribution with the ground plane as a reference, and an interpolation algorithm is used to generate a continuous height curve. The highest points on the left and right sides are located on this curve, and tangents are fitted to obtain three types of geometric feature parameters: tangent angle difference, average height difference, and highest point height difference. Finally, these feature parameters are combined into a feature vector and input into a trained fully connected neural network model. This network contains two hidden layers, each with 128 neurons, using the ReLU activation function, and incorporates Dropout and L2 regularization to suppress overfitting. The network output is the scoliosis angle (Cobb angle), and the degree of spinal curvature is determined based on the range of the output value.

[0020] Further, S1 includes the following steps: S11. Based on the camera pose calibration parameters, obtain the intrinsic and extrinsic parameter matrices of the laser depth camera, perform registration calculations on the camera optical axis direction and the ground normal vector, and obtain the unified transformation matrix between the camera coordinate system and the world coordinate system. S12. Based on the unified transformation matrix, voxel downsampling is performed on the airfield point cloud data, the centroid coordinates of the point cloud within the voxels are calculated and the mean point cloud set is output; RANSAC plane fitting is performed on the mean point cloud set to obtain the ground plane equation and the set of points in the plane. S13. Calculate the vertical height of the camera from the ground based on the ground plane equation and the coordinates of the camera's optical center, and use this height as a scale reference to normalize the point cloud coordinates to obtain spatially normalized point cloud data. S14. Based on the normalized point cloud data and the camera intrinsic parameter matrix, perform three-dimensional back projection and plane intersection calculation on the pixel coordinates to generate a corresponding sample set of pixel points and ground physical coordinates; fit a mapping function according to the sample set to obtain the mapping relationship between pixel coordinates and physical coordinates.

[0021] In some embodiments, the laser depth camera employs structured light ranging, with an installation height of approximately 2.2 meters and the lens axis pointing vertically towards the center of the ground. The system first reads the camera's intrinsic and extrinsic parameter matrices from the camera calibration file. By registering the camera's optical axis direction vector with the ground normal vector, it uses least-squares optimization to obtain the rotation matrix, thus forming a unified spatial transformation matrix and unifying the camera coordinate system with the world coordinate system. Compared to traditional geometric correction methods relying on single-plane calibration, this method can automatically correct camera attitude errors under non-ideal installation conditions, significantly improving the baseline accuracy of subsequent spatial reconstruction. After obtaining the unified transformation matrix, the system performs voxel downsampling on the airfield point cloud data, dividing the point cloud into fixed voxel units and calculating the centroid of the point cloud within each voxel, outputting the mean point cloud set to reduce noise interference. Subsequently, a random sampling consensus algorithm is used for ground plane fitting. Each time, three non-collinear points are randomly sampled and substituted into the plane equation. The distance from all points to the plane is calculated, and the set of interior points is statistically analyzed. Through multiple iterations, a plane with an interior point ratio of no less than 95% is selected as the optimal ground plane, resulting in the ground plane equation and the set of interior points. Based on the optimal ground plane equation and the camera's optical center coordinates, the vertical height from the camera to the ground is calculated. This height is then used as a scale reference to normalize the point cloud spatial coordinates, ensuring all measurement data are expressed at a uniform physical scale. Unlike traditional calibration methods that only utilize pixel scaling factors, this method achieves absolute scale mapping through real physical height, ensuring physical consistency in subsequent height difference and angle calculations. Furthermore, based on the normalized point cloud data and the camera's intrinsic parameter matrix, a 3D back-projection calculation is performed on the pixel coordinates to obtain the intersection coordinates of the pixel ray and the ground plane. The system generates a corresponding sample set of pixel points and ground physical coordinates through multi-frame sampling, and uses polynomial fitting to establish a transformation function between pixel coordinates and physical coordinates, achieving accurate mapping from pixel space to real physical space.

[0022] Further, performing RANSAC plane fitting on the mean point cloud to obtain the ground plane equation and the set of points in the plane includes the following steps: S131. Based on the set of mean points, three points located in different spatial positions and not collinear are randomly selected from them. They are substituted into the plane equation, and the plane parameters of the current round are obtained by solving the system of linear equations. S132. Based on the plane parameters, calculate the Euclidean distance from all points in the mean point cloud to the plane, and set the distance threshold to 0.01m. Define points whose distance does not exceed the threshold as interior points, and count the number of interior points in the current round. S133. Repeat steps S131-S132 500 times, and select the plane with the most interior points and an interior point ratio of not less than 95% as the optimal ground plane. S134. Retain the plane equation and its interior point set corresponding to the optimal ground plane.

[0023] In some embodiments, firstly, three spatially distinct and non-collinear points are randomly selected from the mean point cloud. These points are then substituted into the plane equation to solve for the parameters, resulting in the plane model for the current iteration. This process avoids the bias problem inherent in traditional least-squares fitting methods when outliers are present, making the fitting process more robust. Subsequently, the distance from all points to the plane is calculated based on the current plane model. A distance threshold of 0.01 meters is set, and points with a distance not exceeding this threshold are defined as inliers, and their number is counted. This threshold setting has been verified through extensive experiments, demonstrating that it maintains high fitting accuracy even under conditions of slight ground unevenness and camera depth noise. The random sampling and inlier counting steps are iterated 500 times, recording the number and distribution of inliers in each iteration. Finally, the plane with the highest number of inliers and an inlier percentage of at least 95% is selected as the optimal ground plane. This strategy effectively suppresses the uncertainty caused by random sampling through iterative statistical methods, improving the global stability of the fitting results. Finally, the plane equation corresponding to the optimal ground plane and its inlier set are retained as input data for subsequent calculations of the camera's vertical height from the ground and for constructing a spatial scale benchmark. Compared to traditional ground estimation methods based on single-frame depth maps, this embodiment achieves high-confidence ground modeling through the spatial statistical characteristics of point clouds, significantly improving the accuracy and repeatability of ground identification in complex scenes.

[0024] Furthermore, the step of synchronously acquiring the RGB image and point cloud data of the detection object according to the mapping relationship includes the following steps: Based on the mapping relationship between pixels and physical coordinates, external parameter calibration and time synchronization are performed on the laser depth camera and the RGB camera to establish a unified timestamp and spatial coordinate system and obtain the registration parameters between the cameras. Based on the registration parameters, multiple frames of RGB image data and depth point cloud data of the detected object within the acquisition area are obtained, and the camera pose and exposure parameters are recorded simultaneously during the acquisition process.

[0025] In some embodiments, Furthermore, the YOLOv8-s model is trained through the following steps: The pre-acquired image data is labeled with back and head, and then size normalization, brightness enhancement and rotation perturbation are performed to obtain the enhanced training sample set. Based on the training sample set, a YOLOv8-s detection network model is constructed. Feature extraction and target classification training are performed on the input image. Initial detection weight parameters are output through the cosine annealing learning rate scheduling mechanism and the Adam optimization algorithm. Transfer learning is performed based on the initial detection weight parameters. The backbone network is fixed and only the detection head is trained to obtain the optimized weights of the detection head. Then, the entire network is fine-tuned and trained to output the converged complete model weights. Based on the complete model weights, the model is converted into the OpenVINO intermediate representation format, and the parameters are quantized using FP16 to obtain a lightweight YOLOv8-s model.

[0026] In some embodiments, after establishing the pixel-to-physical coordinate mapping relationship, joint spatial extrinsic parameter calibration and time synchronization are performed on the laser depth camera and the color camera. Spatial extrinsic parameter calibration employs a target-free self-consistent method: using the edge and texture consistency between the airfield point cloud and the color image as constraints, the point cloud is first projected onto a ground reference plane in the physical coordinate system, and then the projection result is projected back onto the color image plane via the aforementioned mapping relationship. By maximizing the overlap between depth edges and brightness gradients, the rotation and translation parameters between the two cameras are iteratively solved, and the reprojection mean square error is used as the convergence criterion. For time synchronization, a hardware triggering and master-slave clock joint strategy is adopted: the depth camera acts as the time master, sending synchronization pulses to the color camera, while simultaneously timestamping all sensor data using a unified time base in the industrial control computer; trigger response jitter and arrival delay are calculated for each batch of synchronization frames, and if they exceed a set threshold, the batch is automatically discarded and reacquired to ensure cross-device time consistency. To further eliminate errors introduced by rolling shutter speed and exposure differences, the system records the exposure time, gain, and white balance parameters of the color camera during acquisition, as well as the projection power and integration time of the depth camera. In post-processing, these parameters are used as conditions for photometric normalization and depth confidence weighting. After completing external participation time synchronization, the system initiates multi-frame synchronous acquisition once the inspected object enters the acquisition area. Typically, it continuously acquires several dozen frames of color images and corresponding point clouds, while recording the camera attitude in real time (estimated through inertial measurement or visual odometry) to eliminate registration errors introduced by minor body movements. Each pair of synchronized frames undergoes spatial registration verification, with verification indicators including reprojection error, point cloud-to-contour overlap rate, and timestamp consistency. Only frame pairs that simultaneously meet the threshold constraints are written into the valid data cache. Unlike existing technologies that separate time synchronization and extrinsic parameter calibration and typically rely on manual targets, this embodiment places spatial registration and time alignment in the same optimization loop and uses the ground reference plane and mapping relationship as physical scale constraints to achieve self-calibration and robust alignment under natural environments and non-ideal attitudes.

[0027] Furthermore, the pre-collected image data includes human sample images of different genders, ages, heights, body types, clothing types, lighting conditions and background environments, and posture deviations.

[0028] It should be noted that the pre-collected image data comes from a multi-center collaborative human sample database, aiming to ensure the diversity of training data and the model's generalization ability. The subjects collected cover different genders, ages, heights, and body types, with a focus on adolescents to capture diverse morphological characteristics during spinal development. The samples include individuals of normal build, thin, and obese physiques to ensure the model's robustness in representing the back contour under different body proportions. Regarding clothing, data collection covers various types and materials of clothing, such as school uniforms, sportswear, sweaters, sweatshirts, and tight-fitting clothes, to simulate occlusion in real-life medical examinations or everyday scenarios; and sample equalization is performed based on clothing color and texture variations to avoid recognition bias caused by specific color distributions. For lighting and background conditions, samples were collected under natural light, LED lighting, backlighting, and low-light conditions, while also incorporating various environments such as classrooms, gymnasiums, medical examination rooms, and outdoors to enhance the model's adaptability to complex backgrounds. Furthermore, different posture deviations were artificially designed during the sampling process, including slight back bending, left and right torso rotation, and forward tilt angle changes. This ensured that the dataset contained multi-angle posture features, guaranteeing that the model could accurately locate the center point of the back and head even in non-standard standing postures. Compared with existing training samples that rely on single body type or ideal posture data, this embodiment, through the introduction of a systematic multi-dimensional acquisition design, achieved coordinated coverage of five dimensions: gender, body type, clothing, lighting, and posture. This fundamentally improved the robustness and generalization ability of the YOLOv8-s detection model in natural scenes.

[0029] Further, S3 includes the following steps: Based on the initial line of the spine center and the mapping relationship between pixels and physical coordinates, the back region of the detected object is spatially projected, and the vertical projection range of the spine center line in the physical coordinate system is divided into chest, thoracolumbar and lumbar regions to obtain the corresponding region boundary coordinates. Based on the boundary coordinates of the region, a left-right symmetrical sampling window is constructed in each segmented region. Three-dimensional back projection calculation is performed on the pixel coordinates in each sampling window to obtain the candidate point cloud datasets on the left and right sides. Based on the point cloud candidate dataset, statistical filtering and outlier removal are performed on the point cloud of each segment region. The point cloud density distribution and spatial variance are calculated, and effective point cloud data after noise suppression are obtained. Based on the effective point cloud data, according to the segmented regions of chest, chest-waist, and waist and their left-right symmetry, a point cloud subset with uniform local spatial features is extracted and combined to obtain the key region point cloud dataset.

[0030] In some embodiments, the back region of the detected object is first mapped to a unified physical coordinate system. Using the initial line of the spine's center as the geometric reference, the physical projection range of this line in the vertical direction is calculated. Based on the anatomical proportions of the human body and the distribution characteristics of the physiological curvature of the spine, this range is divided into three parts: thoracic, thoracolumbar, and lumbar segments, and corresponding region boundary coordinates are automatically generated. This segmentation method differs from traditional partitioning methods based on image height ratios; it achieves physical consistency in region division through a real spatial scale, enabling subsequent feature extraction to accurately reflect the curvature changes of different spinal segments. Within each segmented region, symmetrical sampling windows are constructed based on the boundary coordinates. The width of the windows is adaptively adjusted according to the standard deviation of the individual's back point cloud distribution to accommodate different body types. Three-dimensional back-projection calculations are performed on the pixel coordinates within the windows, mapping two-dimensional pixels to three-dimensional spatial points, thereby generating candidate point cloud datasets on the left and right sides. To ensure the integrity of the point cloud structure, nearest neighbor interpolation is used during the projection process to fill in sparse areas, making the spatial point distribution more continuous. Statistical filtering and outlier removal are then performed on the candidate point cloud datasets. The statistical filtering stage calculates the mean and standard deviation based on the distance distribution of neighboring points. Points deviating from the neighborhood mean by more than twice the standard deviation are considered noise points and removed. The outlier removal stage combines point cloud density estimation to remove isolated points whose density within a local voxel is less than half the global average density. This dual filtering significantly reduces noise interference caused by clothing wrinkles, light reflection, and depth measurement errors. Subsequently, the spatial variance and point density distribution of the filtered point cloud are analyzed, serving as the basis for selecting uniform region point cloud subsets. For each segmented region, a subset with uniform local spatial features is extracted from the effective point clouds of both left and right symmetries. The morphological comparability of the left and right symmetric regions is ensured by Euclidean distance and normal vector consistency constraints. Finally, the system merges the left and right point cloud subsets of each segmented region to form a complete key region point cloud dataset.

[0031] Further, S4 includes the following steps: Based on the point cloud data of the key area and the ground plane equation, the vertical distance from the point to the plane is calculated for the left and right point clouds of each segment area to obtain the corresponding height distribution dataset; Based on the height distribution dataset, the height values ​​of the left and right point clouds in each region are sorted and linearly interpolated according to the horizontal index to generate continuous left and right height curves. Based on the left and right height curves, locate the highest left and right points in each region, and extract local point cloud data in the neighborhood of the highest point; Based on the local point cloud data, the least squares method is used to fit the tangents within ±5mm of the highest point to obtain the left and right tangent angle values. Based on the left and right tangent angle values ​​and the left and right height curves, calculate the tangent angle difference, average height difference, and highest point height difference for each region.

[0032] In some embodiments, based on key region point cloud data and ground plane equations, the vertical distances from points to the ground plane are first calculated for the left and right point clouds of three segmented regions: chest, thoracolumbar, and lumbar, thereby obtaining a height distribution dataset reflecting the undulations of the back contour. This step uses the fitted ground equation as a unified benchmark to ensure the comparability of height quantification results for different individuals and camera mounting postures, avoiding measurement biases caused by posture shifts in traditional two-dimensional methods based on relative pixel differences. Subsequently, the left and right height distribution data of each region are reordered and linearly interpolated according to the horizontal index to generate continuous height curves at 0.5 mm intervals, reconstructing a smooth surface contour from the discrete point cloud. This process effectively fills in data gaps caused by occlusion, reflection, or sparse point clouds, providing a continuous geometric basis for subsequent curvature and angle analysis. On the generated left and right height curves, the highest points of each region are located by searching for local extrema, serving as feature points for the back spine shape. To avoid false peaks caused by clothing wrinkles or hair interference, if a highest point is only higher than a single neighboring point or forms an isolated spike, the landing point position is re-determined through neighborhood smoothing and second-highest point replacement algorithms. Ultimately, each region yields a set of left and right highest points with a spatial positioning error not exceeding 0.5mm, perfectly corresponding to the contact logic of a clinical "spine ruler" for detecting the human back. Based on the located highest points, the system extracts an interpolation point set within its local neighborhood (±5mm range), performs linear fitting using the least squares method, solves the tangent equation, and calculates its angle with the horizontal direction. If the sum of squared fitting residuals exceeds the threshold of 0.01mm², the fitting neighborhood is automatically expanded to ±8mm for recalculation, ensuring the tangent accurately reflects the tilting trend of the back protrusion. The arctangent operation yields the tangent angle values ​​on both sides, with the calculation error controlled within 0.2°. Finally, combining the left and right tangent angle values ​​and the height curve, the tangent angle difference, average height difference, and highest point height difference for each segment are calculated, forming a geometric feature vector describing the differences in the three-dimensional morphology of the back. This process breaks through the limitations of traditional two-dimensional image planar analysis in terms of methodology, and realizes three-dimensional structural quantification based on physical spatial parameters. It maintains comparability with clinical physical measurements (spinal ruler landing point and tilt angle) and significantly improves the spatial accuracy and interpretability of automatic detection.

[0033] Further, S5 includes the following steps: Based on the data of tangent angle difference, average height difference and highest point height difference, multidimensional feature vectors are constructed according to the chest, chest-waist and waist regions. Each feature is input into a trained fully connected neural network model, and multi-layer linear mapping and non-linear activation calculations are performed on the input data to obtain the predicted values ​​of spinal curvature in each region. The global scoliosis angle is calculated based on the predicted spinal curvature value, and the prediction results of different regions are fused by weighted averaging to output the overall spinal scoliosis angle. The spinal curvature level is determined and output by matching the overall scoliosis angle with a preset level threshold table.

[0034] In some embodiments, the tangent angle difference, average height difference, and highest point height difference extracted in the previous stage are used as the main input features, and a multi-dimensional feature vector is constructed based on the spatial distribution of the thoracic, thoracolumbar, and lumbar segments. This feature structure not only preserves the independent morphological features of different anatomical regions of the back, but also achieves numerical comparability across regions through a unified vectorized representation. The multi-dimensional feature vector is input to a trained fully connected neural network model, which contains two hidden layers, each with 128 neurons. The activation function is ReLU, and a Dropout mechanism is introduced between each layer to suppress overfitting. Model training uses the Cobb angle measured by X-rays as the supervision label, and iterative training is performed using the mean squared error loss function and the Adam optimization algorithm to achieve an end-to-end mapping from the multi-feature space to the spinal curvature angle. Unlike traditional linear regression methods based on a single geometric parameter, this embodiment achieves the learning and fitting of complex coupling relationships between multi-dimensional features through multi-layer mapping and nonlinear activation of deep neural structures. In the prediction stage, the model outputs predicted values ​​of spinal curvature for the thoracic, thoracolumbar, and lumbar segments, respectively. A weighted average is calculated based on the weight differences of feature regions (determined by feature importance analysis) and fused into the overall scoliosis angle. This fusion method takes into account the different contributions of different regions to the global curvature morphology, ensuring that even mild and occult curvatures can be accurately quantified. Finally, the system performs interval matching between the output overall spinal scoliosis angle and a preset threshold table to automatically determine the curvature level, including four categories: normal, mild, moderate, and severe. This automatic grading strategy achieves a closed-loop recognition process from physical spatial geometric features to clinically determined grades. Compared to existing technologies that require manual interpretation or rely on X-ray images, this embodiment achieves clinical-level angle quantification and grading accuracy at a low cost.

[0035] Furthermore, the fully connected neural network is constructed through the following steps: A training sample set is established based on the multidimensional feature vector and the corresponding actual Cobb angle data of the spine, and the input features are normalized to obtain standardized feature data. Based on the standardized feature data, the number of neurons in the input layer is set to be consistent with the feature dimension, and a fully connected network structure containing two hidden layers is constructed. Each hidden layer is activated by the ReLU function and Dropout is set to suppress overfitting to obtain the initial network model. Based on the initial network model and training sample set, iterative training is performed using the mean squared error loss function and the Adam optimization algorithm. Gradient updates are performed on the network weight parameters until the loss converges, and the converged model parameters are obtained. Based on the converged model parameters, a prediction calculation is performed on the validation data, the predicted scoliosis angle is output, and the network parameters that meet the accuracy threshold are solidified into a fully connected neural network that has been trained.

[0036] In some embodiments, multidimensional geometric features (tangent angle difference, average height difference, and highest point height difference) from three key segments (thoracic, thoracolumbar, and lumbar) are first concatenated in a fixed order to form a sample feature matrix. To eliminate the influence of differences in dimensions and magnitudes on the optimization process, all input features are standardized (using training set statistics as a benchmark for unified validation and testing), and outliers are gently truncated to ensure that extreme samples do not dominate parameter updates. To ensure the reliability of the labels, angle labels are averaged after independent measurements by two orthopedic surgeons, and divergent samples are reviewed to establish a training sample set and an independent validation set (divided by subject to avoid information leakage caused by the same subject crossing sets). Based on the standardized feature matrix, a fully connected regression network is constructed: the input layer dimension corresponds one-to-one with the feature dimension; the two hidden layers use the same size to enhance representational power, and the activation function uses a modified linear unit to obtain a stable gradient; random deactivation and weight decay are added after each hidden layer to suppress overfitting; the weight initialization adopts a variance adaptive strategy to ensure that the initial distribution matches the activation function. The architecture is designed with the following motivations: compared to traditional linear regression or support vector regression, multi-layer mapping can learn the non-linear coupling relationships between segments and features; compared to convolutional or temporal structures, fully connected layers are more suitable for the fixed-length, low-dimensional, and statistical inputs of this scenario, with lower inference costs and easier deployment on resource-constrained terminals. During the training phase, mean squared error is used as the optimization objective, and an adaptive moment estimation algorithm is used for parameter updates, coupled with cosine annealing-based learning rate scheduling and an early stopping strategy: training stops and rolls back to the optimal weights when the validation set loss does not improve within several rounds. To further improve generalization ability, small noise perturbations and order perturbations are applied to the features during training, and key hyperparameters are optimized using grid or Bayesian methods with hold-out or cross-validation. After training, performance is evaluated using an independent validation set, reporting the mean absolute error, root mean square error, and error distribution under different severity levels; simultaneously, calibration curves and coefficients of determination are calculated to verify the numerical consistency and explanatory power of the model output. If the verification metrics meet the preset threshold, the obtained weights are solidified into an inference model, and standardized parameters and post-processing procedures are recorded to form a reproducible experimental package to support batch deployment.

[0037] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for identifying abnormal curvature of the spine in the back of a human body based on a depth camera, characterized by, The method comprises the following steps: S1, fixing a laser depth camera according to preset installation parameters, collecting empty field point cloud data, performing voxel downsampling on the point cloud, fitting a ground plane using a RANSAC algorithm, calculating the vertical height of the camera to the ground according to the fitted equation, and obtaining the mapping relationship between pixels and physical coordinates; S2, according to the mapping relationship, synchronously collecting the RGB image and the point cloud data of the detection object, inputting the image into a YOLOv8-s model to identify the back and head center points, and calculating the connecting line according to the coordinates of the two points to obtain an initial line of the spine center; S3, based on the initial line of the spine center, dividing the back into three segments of chest, thoracolumbar, and lumbar, determining left and right symmetrical sampling windows according to the pixel-physical mapping, extracting corresponding point cloud subsets and performing filtering processing to obtain key region point cloud data; S4, according to the key region point cloud data, calculating the left and right height distribution with the ground plane as the reference, performing interpolation on the height values to generate a height curve, positioning the left and right highest points in the curve and fitting a tangent line to obtain the tangent angle difference, average height difference and highest point height difference; S5, inputting the characteristic vectors of the tangent angle difference, average height difference and highest point height difference into a fully connected neural network model to output the scoliosis angle of the spine, and determining the curvature grade of the spine according to the belonging preset range of the scoliosis angle. 2.The method of claim 1, wherein, The S1 comprises the following steps: S11, according to the camera pose calibration parameters, obtaining the intrinsic matrix and extrinsic matrix of the laser depth camera, performing registration calculation on the camera optical axis direction and the ground normal vector to obtain the unified transformation matrix of the camera coordinate system and the world coordinate system; S12, based on the unified transformation matrix, performing voxel downsampling on the empty field point cloud data, calculating the centroid coordinates of the point cloud in the voxel and outputting a mean point cloud set; performing RANSAC plane fitting on the mean point cloud set to obtain a ground plane equation and a set of points in the plane; S13, calculating the vertical height of the camera to the ground according to the ground plane equation and the optical center coordinates of the camera, and performing scale normalization on the point cloud coordinates with the height as the scale reference to obtain space scale normalized point cloud data; S14, based on the normalized point cloud data and the camera intrinsic matrix, performing three-dimensional back projection and plane intersection calculation on the pixel coordinates to generate a corresponding sample set of pixel points and ground physical coordinates; fitting a mapping function according to the sample set to obtain the mapping relationship between pixel coordinates and physical coordinates. 3.The method of claim 2, wherein, The RANSAC plane fitting on the mean point cloud set to obtain the ground plane equation and the set of points in the plane comprises the following steps: S131, according to the mean point cloud set, randomly selecting three points at different spatial positions and not collinear from the mean point cloud set, substituting them into the plane equation, and solving the plane parameters of the current round through solving a linear equation set; S132, based on the plane parameters, calculating the Euclidean distance of all points in the mean point cloud set to the plane, setting the distance threshold to 0.01 m, defining the points with a distance not exceeding the threshold as inliers, and counting the number of inliers in the current round; S133, repeat the steps S131-S132 for 500 times, and select a plane with the largest number of inliers and an inlier ratio of no less than 95% as an optimal ground plane; S134, reserve the plane equation corresponding to the optimal ground plane and the inlier set thereof. 4.The method of claim 1, wherein, The step of synchronously collecting the RGB image and the point cloud data of the detection object according to the mapping relationship comprises the following steps: Based on the mapping relationship between the pixels and the physical coordinates, the extrinsic calibration and time synchronization of the laser depth camera and the RGB camera are performed to establish a unified timestamp and spatial coordinate system, and the registration parameters between the cameras are obtained; According to the registration parameters, a plurality of frames of RGB image data and depth point cloud data of the detection object in the collection region are obtained, and the camera pose and exposure parameters are recorded synchronously during the collection process. 5.The method of identifying abnormal curvature of a human back spine based on a depth camera according to claim 1, characterized in that, The YOLOv8-s model is trained by the following steps: The pre-collected image data is labeled for back and head, and size normalization, brightness enhancement and rotation disturbance processing are performed to obtain an enhanced training sample set; Based on the training sample set, a YOLOv8-s detection network model is constructed to perform feature extraction and target classification training on the input image, and an initial detection weight parameter is output through a cosine annealing learning rate scheduling mechanism and an Adam optimization algorithm; According to the initial detection weight parameter, transfer learning is performed, the backbone network is fixed, and only the detection head is trained to obtain an optimized weight of the detection head; Then, the full network is fine-tuned to output a converged complete model weight; Based on the complete model weight, the model is converted to an OpenVINO intermediate representation format, and FP16 quantization processing is performed on the parameters to obtain a lightweight YOLOv8-s model. 6.The method of claim 5, wherein, The pre-collected image data includes human sample images of different genders, ages, heights, body types, clothing types, lighting conditions, background environments and posture deviations. 7.The method of claim 1, wherein, The S3 comprises the following steps: Based on the initial spine center line and the mapping relationship between the pixels and the physical coordinates, the back region of the detection object is spatially projected, and the vertical projection range of the spine center line in the physical coordinate system is divided into chest, thoracolumbar and lumbar to obtain corresponding regional boundary coordinates; According to the regional boundary coordinates, left-right symmetric sampling windows are constructed in each segmented region, and three-dimensional back projection calculation is performed on the pixel coordinates in each sampling window to obtain left and right point cloud candidate data sets; Based on the point cloud candidate data set, statistical filtering and outlier rejection processing are performed on the point cloud of each segmented region, the point cloud density distribution and spatial variance are calculated, and the effective point cloud data after noise suppression are screened. Based on the effective point cloud data, point cloud subsets with uniform local spatial features are extracted according to the chest, thoracolumbar and lumbar segmented regions and their left-right symmetry relationship, and a key region point cloud data set is obtained. 8.The method of claim 1, wherein, The S4 comprises the following steps: Based on the key region point cloud data and the ground plane equation, the vertical distance from the points to the plane is calculated for the left and right point clouds of each segmented region to obtain a corresponding height distribution data set; According to the height distribution data set, the height values of the left and right point clouds of each region are sorted and linearly interpolated according to the horizontal direction index to generate continuous left and right height curves; Based on the left and right height curves, the left and right highest points in each region are located, and the local point cloud data in the neighborhood of the highest points is extracted; According to the local point cloud data, the least square method is used to fit the tangent line of the data within ±5mm range of the highest point to obtain the left and right tangent angle values; According to the left and right tangent angle values and the left and right height curves, the tangent angle difference, the average height difference and the highest point height difference of each region are calculated. 9.The method of claim 1, wherein, S5 includes the following steps: Based on the tangent angle difference, the average height difference and the highest point height difference data, a multi-dimensional feature vector is constructed according to the chest, thoracolumbar and lumbar regions; Each feature is input into a trained fully connected neural network model, and multi-layer linear mapping and nonlinear activation calculation are performed on the input data to obtain the prediction value of the curvature of each region of the spine; Based on the prediction value of the curvature of the spine, the global scoliosis angle is calculated, the prediction results of different regions are fused by weighted average, and the overall scoliosis angle of the spine is output; According to the overall scoliosis angle of the spine and the preset grade threshold table, interval matching is performed, the curvature grade of the spine is determined, and the result is output. 10.The method of claim 9, wherein, The fully connected neural network is constructed by the following steps: According to the multi-dimensional feature vector and the corresponding actual Cobb angle data of the spine, a training sample set is established, and the input features are normalized to obtain standardized feature data; Based on the standardized feature data, the number of input layer neurons and the feature dimension are set to be consistent, a fully connected network structure containing two hidden layers is constructed, each hidden layer is activated by ReLU function and set to suppress overfitting by Dropout, and an initial network model is obtained; According to the initial network model and the training sample set, the mean square error loss function and the Adam optimization algorithm are used for iterative training, the gradient of the network weight parameters is updated until the loss converges, and the converged model parameters are obtained; Based on the converged model parameters, prediction calculation is performed on the verification data, the scoliosis angle prediction result is output, and the network parameters meeting the accuracy threshold are solidified as the trained fully connected neural network.

Images