A non-contact automatic height measurement system based on multi-view image fusion

Abstract

The application provides a non-contact automatic height measurement system based on multi-view image fusion, which synchronously collects front and side images of a human body through double cameras, automatically guides the user to stand, evaluates the posture stability in real time, and after image distortion removal, semantic segmentation and key point extraction, uses the polar line constraint and SuperGlue matching algorithm to associate the features across the view angles, and through triangulation and bundle adjustment optimization, obtains millimeter-level three-dimensional coordinates. The posture tilt cosine correction and ground normal vector compensation model are innovatively introduced, which effectively eliminates the errors caused by body forward leaning or uneven ground. The system has automatic calibration, privacy protection and remote operation and maintenance functions. The application can be widely applied to hospitals, schools and community scenes, realizes unmanned, safe and sanitary, efficient and accurate automatic height monitoring, and has significant social benefits and industrialization value.

Description

Technical Field

[0001] This invention relates to the field of height measurement, and more specifically, to a non-contact automatic height measurement system based on multi-view image fusion. Background Technology

[0002] Height is an important physiological indicator for measuring human growth, development, nutritional status, and health level, and is widely used in clinical medicine, public health, sports training, and health management. Traditional height measurement mainly relies on contact measuring tools, such as mechanical height rulers and wall-mounted rulers with right-angle headplates. In recent years, with the development of computer vision, artificial intelligence, and 3D sensing technologies, non-contact automatic height measurement systems have gradually become a research hotspot and have begun to be applied in some intelligent physical examination equipment.

[0003] Currently, common non-contact height measurement technologies mainly include the following categories: Monocular vision measurement method: This method uses a single camera combined with a deep learning model to extract key points of the human body, and calculates height through pixel ratio conversion or preset calibration relationships. This method is simple in structure and low in cost, but it is easily affected by factors such as shooting distance, viewing angle tilt, and lens distortion, resulting in unstable measurement accuracy, especially with larger errors in long-distance or non-directly facing scenarios.

[0004] Structured light / ToF depth camera solution: This method uses infrared projection and receiving devices to acquire human depth maps and directly extracts body contours for height calculation. While this approach provides 3D information and has a fast response time, its performance degrades significantly in strong light environments, and its resolution is limited, making it difficult to accurately identify the top and bottom of the head. It often results in head loss or ground blurring issues, leading to a generally low systemic performance.

[0005] The Kinect RGB-D sensor integrated system uses Microsoft Kinect or other multimodal sensors to achieve human skeleton tracking and height estimation. Although it has a certain 3D perception capability, its effective working distance is short (usually ≤3m), the field of view is small, and it has strict requirements on human posture (the human must stand facing the device). It performs poorly in scenarios where multiple people pass by quickly or in complex standing postures.

[0006] Laser scanning height measurement system: This system uses a rotating lidar to vertically scan the human body's outline and calculates height by fitting the intersection of the highest point and the ground. While this type of system offers high accuracy, the equipment is expensive, bulky, and consumes a lot of power, making it unsuitable for large-scale deployment in community and school settings.

[0007] Despite the progress made in the aforementioned non-contact measurement technologies, many key technical bottlenecks and practical shortcomings still exist in their application, as follows: Existing systems lack a rigorous three-dimensional geometric reconstruction mechanism, relying solely on two-dimensional image scaling or coarse depth estimation, leading to significant fluctuations in measurement results. For example, the magnification variation of monocular vision methods at different distances is not effectively compensated, resulting in height errors exceeding ±1.5cm, making them unsuitable for clinical diagnosis or long-term health monitoring.

[0008] Existing systems generally require the subject to face the device directly, with feet together and head held high. Even a slight forward tilt, sideways movement, or head turn can lead to key point localization failure or projected shortening, resulting in significant negative bias. The lack of effective posture correction algorithms limits their applicability in real-world, complex environments.

[0009] Most systems assume the ground is an ideal level surface and do not consider the slight slopes that may exist in the actual installation environment, such as uneven carpets or floors. When the subject stands on an inclined surface, the reference plane of the foot shifts, causing inaccurate height calculation benchmarks and introducing systematic errors.

[0010] Image-based measurement methods require capturing complete human appearance data. If the raw images are stored or transmitted without processing, they can easily reveal personal identification features. The existing system still requires operators to manually trigger shooting, confirm station positions, or remove abnormal data, thus failing to achieve true unattended operation. Furthermore, after long-term operation, camera parameter deviations caused by temperature drift and vibration are not automatically detected and corrected, leading to decreased measurement consistency.

[0011] When the subject's arms are hanging down and obscuring their torso, when they are wearing loose clothing, or when multiple people are standing close together, key point detection is prone to errors or even failure. The system often directly outputs incorrect results or interrupts the measurement, lacking intelligent repair and anomaly judgment mechanisms.

[0012] Traditional multi-view systems require periodic recalibration of external parameters using standard rods or calibration plates, a complex process dependent on specialized technicians, making them unsuitable for widespread application in grassroots units or remote sites. Therefore, we propose an improvement: a non-contact automatic height measurement system based on multi-view image fusion. Summary of the Invention

[0013] The purpose of this invention is to address the problems raised in the existing background technology. To achieve the above-mentioned objective, this invention provides the following technical solution: a non-contact automatic height measurement system based on multi-view image fusion, comprising a multi-view image acquisition module, equipped with at least two wide-angle CMOS cameras with a resolution of 1920×1080 and a frame rate of 30fps, installed on the left and right sides of the measurement area, with a horizontal angle of 45°±3°, a height of 1.5m±0.05m, a spacing of 1.2m, and a lens field of view ≥85° (H)×55° (V), supporting hardware-level synchronous triggering (time deviation ≤1ms), used to acquire frontal and side images of the subject; Image preprocessing module: Receives the raw images transmitted from the multi-view image acquisition module, performs denoising processing using an adaptive median filtering algorithm, with the filter window size dynamically adjusted based on the noise density of local image regions. Image enhancement is then performed using the Adaptive Histogram Equalization (CLAHE) method, with a contrast enhancement constraint factor of [missing information]. The value range is 1.0-4.0. Distortion correction is performed using pre-calibrated camera distortion parameters. The image pixel coordinates are determined according to the following formula. Perform correction: in, Multi-view image fusion module: This module uses a feature matching algorithm to extract and match features from preprocessed multi-view images, then fuses the matched images to produce a fused image containing complete human information. The fusion process employs a Laplacian pyramid-based fusion method, and the pixel values ​​of the fused image are calculated using the following formula. : in, These are the pixel values ​​of the two images to be merged. and The weights are calculated based on the image's sharpness and credibility. ; The 3D spatial reconstruction module is based on a pre-calibrated intrinsic parameter matrix. and external references ,use The algorithm initializes the 3D coordinates and achieves this by minimizing the reprojection error. optimization; Height calculation module: Combines pre-calibrated system parameters (such as the camera's focal length) Based on the installation location and angle, as well as the extracted human feature point information, the actual height of the human body is calculated using the principle of triangulation. Assume the angle between the camera's optical axis and the horizontal direction is... The vertical pixel distance between the head vertex and the foot point on the image plane is Calculate a person's actual height using the following formula: At the same time, taking into account the physiological curvature of the human body, the calculation results are corrected, with a correction factor of [value missing]. The value range is 0.99-1.01, which is the corrected height. ; The data fusion and result output module performs a moving average filter on three consecutive valid measurement values ​​and finally outputs the result with an accuracy of 0.1cm. System calibration module: Performs periodic system calibration to ensure measurement accuracy. Standard height used is... The system is calibrated using a reference object with a height accuracy of ±0.05 cm. During calibration, various system parameters, such as the camera's focal length, are remeasured and updated. Installation location and angle. Calculate the calibrated focal length using the following formula. : in, This is the vertical pixel distance of the reference object on the image plane. The calibration cycle is every two weeks or can be adjusted based on actual usage. Pose detection and correction module: By analyzing and fusing the pose information of the human body in the image, it determines whether the human body is in an upright position. The OpenPose algorithm is used to detect the joint angles of the human body. When the system detects that the human posture does not meet the requirements for standing, it issues a voice prompt to guide the subject to adjust their posture until the correct measurement posture is achieved. The cross-device data consistency calibration module automatically verifies system accuracy and triggers recalibration daily using a standard mannequin.

[0014] Data transmission module: Supports the transmission of measured height data to other devices or systems via wired networks (such as Ethernet) or wireless networks (such as 5G), with a transmission rate of no less than 5Mbps, to achieve data sharing and remote monitoring. Display module: Uses an LCD screen to display the measured height data, measurement time, and accuracy information of the measurement results. The screen size is no less than 10 inches, and the resolution is no less than 1280×800 pixels, for user convenience in viewing and operation. As a preferred technical solution of the present invention, any three-dimensional key point in the three-dimensional space reconstruction module In the Projected coordinates on the imaging plane of each camera Satisfying the pinhole camera projection model: in As a scale factor, For the first Intrinsic parameter matrix of each camera: As a preferred technical solution of the present invention, the three-dimensional space reconstruction module further employs a nonlinear optimization method—beam adjustment. The following objective function is constructed to minimize the reprojection error under all views: in This represents the projection function with distortion correction. Visibility weight (if the first) Point at (If visible from the viewpoint, then it is 1). Optimization variables include the 3D coordinates of all keypoints. The camera pose parameters are solved using the Levenberg-Marquardt algorithm, which brings the root mean square error (RMSE) of the reprojection error to ≤1.8 pixels.

[0015] As a preferred technical solution of the present invention, the posture correction unit in the height calculation module calculates the torso direction vector. and vertical direction The included angle Achieving cosine compensation: The final height correction formula is: like The system prompts you to remain upright and pause the measurement.

[0016] As a preferred technical solution of the present invention, the height calculation module further includes a ground normal vector estimation function: by estimating the ground normal vector of multiple three-dimensional points in the foot area. Perform plane fitting to solve the least squares problem: Obtain the optimal plane normal vector And adjust the height measurement direction from the Z-axis to along Direction, actual height calculated as follows: Suitable for applications where the ground tilt angle is ≤3°.

[0017] As a preferred technical solution of the present invention, the multi-view keypoint matching module employs a basic matrix. Establish epipolar constraints between the two viewpoints: for the left viewpoint point Point corresponding to the right view ,satisfy: in Let be the antisymmetric matrix of the translation vector. The RANSAC algorithm uses this constraint to select interior points and sets a reprojection error threshold. Pixels, maximum number of iterations This ensures robust matching.

[0018] As a preferred technical solution of the present invention, the dynamic attitude stability evaluation module defines key point motion stability indices. For the past Standard deviation of the location of a key point within a frame: When the key points of the hip or ankle If the attitude is deemed unstable, the system delays measurement until all key points have been measured for a continuous period of 0.5 seconds. .

[0019] As a preferred technical solution of the present invention, the anomaly detection and fault tolerance mechanism enables a mirror interpolation repair strategy when the single-view keypoint missing rate exceeds 40%: assuming the left keypoint... If missing, then use the right-side symmetrical point. Normal vector to the central plane of the human body (formed by the tip of the nose, base of the neck, and mid-hip) Calculate the mirror point: in This is the nearest projection point on the central plane. After repair, the reasonableness is verified by combining the prior bone length (e.g., shoulder width ≈ 0.25 × height). If the error exceeds ±15mm, it is marked as low confidence and manual review is prompted.

[0020] A non-contact height measurement method based on a multi-view image fusion non-contact automatic height measurement system includes the following steps: (a) The user guidance process is initiated by detecting the entry of a human body through infrared + ToF composite sensing; (b) Continuously monitor attitude stability indicators ,satisfy Synchronous shooting is triggered after lasting for more than 0.5 seconds; (c) Perform image distortion correction, using the Brown-Conrady model to correct radial and tangential distortions: in $, typical distortion coefficient ; (d) Use HRNet to extract key points, and SuperGlue+RANSAC to complete the matching; (e) The three-dimensional coordinates are reconstructed using a combination of EPNP and Bundling Adjustment. (f) Calculation compensate; (g) Activate the anomaly repair mechanism; (h) Collect 3 frames continuously, remove outliers with a deviation > 0.8 cm, and take the mean; (i) Perform privacy processing: The original image is Gaussian blurred (kernel function) Post-overwrite storage; (j) The results are uploaded in encrypted form, in compliance with GDPR and the Personal Information Protection Act.

[0021] As a preferred embodiment of the present invention, the method is tested on 100 subjects aged 6-80 years and with a height of 100-200 cm under conditions of temperature 15-30℃, relative humidity 30%-70%, and light intensity 300-1000 lx. The systematic mean absolute error (MAE) is: Repeated measures standard deviation coefficient of variation The total time for a single process It also meets all the technical requirements for accuracy, stability and privacy security in the medical device standard YY / T 0654-2020.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: This system does not require the subject to touch any mechanical parts (such as the headboard or sliding ruler). It calculates height entirely through multi-view visual imaging, avoiding the risk of cross-infection caused by multiple people using it continuously. It is especially suitable for public health scenarios such as hospitals, schools, and physical examination centers. The measurement accuracy meets medical-grade standards: within the height range of 100-200 cm, the system mean absolute error (MAE) ≤ 0.45 cm, and the coefficient of variation (CV) for repeated measurements ≤ 0.25%, meeting the requirements of the industry standard "YY / T 0654-2020 Medical Height and Weight Measuring Instrument"; By combining 3D reconstruction and posture compensation algorithms, measurement deviations caused by slight forward tilting, backward tilting, or lateral bending of the human body can be effectively eliminated.

[0023] The system has a complete closed-loop process of detection, guidance, data acquisition, calculation, and output. The entire process from user entry to result display takes no more than 3 seconds and supports continuous measurement of more than 60 people per hour, greatly improving the efficiency of physical examinations. It integrates interactive modules such as voice prompts, laser projection positioning, and LED status indicators, enabling self-service operation with zero training. The elderly and children can complete the operation independently. The dynamic posture stability assessment mechanism ensures that the measurement is triggered only when the human body is stationary, avoiding data errors caused by mis-collection.

[0024] Anomaly detection and fault tolerance mechanisms can enable mirror interpolation repair when key points are occluded, with repair errors controlled within ±15mm, ensuring the success rate of measurements under complex postures.

[0025] The original image undergoes irreversible processing immediately after key point extraction (Gaussian blur kernel σ=15 or pixelation to 32×32), and the cache in memory is destroyed in real time; the system only retains the three-dimensional coordinates of the key points for calculation, and does not store sensitive information such as face and body shape; This invention incorporates a cross-device data consistency calibration module, which automatically calls a standard height simulation dummy (height 150.0±0.1cm) for benchmark testing daily. If the measurement deviation exceeds ±0.3cm, an alarm is triggered and an automatic calibration process is initiated, using a checkerboard calibration board to update internal and external parameters. It supports remote firmware upgrades (OTA), operation log uploads, and fault diagnosis, and is suitable for large-scale distributed deployment. Attached image description: Figure 1 The system module data block diagram provided by this invention; Figure 2 This is a data block diagram of the image acquisition module provided by the present invention; Figure 3 This invention provides an image preprocessing data flowchart; Figure 4 A data flow diagram for maintaining consistency provided by this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 specific implementations of the present invention and are not limited to all embodiments.

[0027] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] It should be noted that, in the absence of conflict, the embodiments and features and technical solutions in the embodiments of the present invention can be combined with each other. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0029] Example: A non-contact automatic height measurement system based on multi-view image fusion includes a multi-view image acquisition module, equipped with at least two wide-angle CMOS cameras with a resolution of 1920×1080 and a frame rate of 30fps, installed on the left and right sides of the measurement area, with a horizontal angle of 45°±3°, a height of 1.5m±0.05m, a spacing of 1.2m, and a lens field of view ≥85° (H)×55° (V), supporting hardware-level synchronous triggering (time deviation ≤1ms), used to acquire frontal and side images of the subject; Image preprocessing module: Receives the raw images transmitted from the multi-view image acquisition module, performs denoising processing using an adaptive median filtering algorithm, with the filter window size dynamically adjusted based on the noise density of local image regions. Image enhancement is then performed using the Adaptive Histogram Equalization (CLAHE) method, with a contrast enhancement constraint factor of [missing information]. The value range is 1.0-4.0. Distortion correction is performed using pre-calibrated camera distortion parameters. The image pixel coordinates are determined according to the following formula. Perform correction: in, Multi-view image fusion module: This module uses feature matching algorithms (such as the ORB algorithm) to extract and match features from preprocessed multi-view images, then fuses the matched images to produce a fused image containing complete human information. The fusion process employs a Laplacian pyramid-based fusion method, calculating the pixel values ​​of the fused image using the following formula. : in, These are the pixel values ​​of the two images to be merged. and The weights are calculated based on the image's sharpness and credibility. ; The 3D spatial reconstruction module is based on a pre-calibrated intrinsic parameter matrix. and external references ,use The algorithm initializes the 3D coordinates and achieves this by minimizing the reprojection error. optimization; Height calculation module: Combines pre-calibrated system parameters (such as the camera's focal length) Based on the installation location and angle, as well as the extracted human feature point information, the actual height of the human body is calculated using the principle of triangulation. Assume the angle between the camera's optical axis and the horizontal direction is... The vertical pixel distance between the head vertex and the foot point on the image plane is Calculate a person's actual height using the following formula: At the same time, taking into account the physiological curvature of the human body, the calculation results are corrected, with a correction factor of [value missing]. The value range is 0.99-1.01, which is the corrected height. ; The data fusion and result output module performs a moving average filter on three consecutive valid measurement values ​​and finally outputs the result with an accuracy of 0.1cm. System calibration module: Performs periodic system calibration to ensure measurement accuracy. Standard height used is... The system is calibrated using a reference object with a height accuracy of ±0.05 cm. During calibration, various system parameters, such as the camera's focal length, are remeasured and updated. Installation location and angle. Calculate the calibrated focal length using the following formula. : in, This is the vertical pixel distance of the reference object on the image plane. The calibration cycle is every two weeks or can be adjusted based on actual usage. Pose detection and correction module: By analyzing and fusing the pose information of the human body in the image, it determines whether the human body is in an upright position. The OpenPose algorithm is used to detect the joint angles of the human body. When the system detects that the human posture does not meet the requirements for standing, it issues a voice prompt to guide the subject to adjust their posture until the correct measurement posture is achieved. The cross-device data consistency calibration module automatically verifies system accuracy and triggers recalibration daily using a standard mannequin.

[0030] Data transmission module: Supports the transmission of measured height data to other devices or systems via wired networks (such as Ethernet) or wireless networks (such as 5G), with a transmission rate of no less than 5Mbps, to achieve data sharing and remote monitoring. Display module: Uses an LCD screen to display the measured height data, measurement time, and accuracy information of the measurement results. The screen size is no less than 10 inches, and the resolution is no less than 1280×800 pixels, for user convenience in viewing and operation.

[0031] In the 3D space reconstruction module, any 3D key point In the Projected coordinates on the imaging plane of each camera Satisfying the pinhole camera projection model: in As a scale factor, For the first Intrinsic parameter matrix of each camera: The 3D space reconstruction module further employs a nonlinear optimization method to construct the following objective function to minimize the reprojection error under all viewpoints: in This represents the projection function with distortion correction. Visibility weight (if the first) Point at (If visible from the viewpoint, then it is 1). Optimization variables include the 3D coordinates of all keypoints. The camera pose parameters are solved using the Levenberg-Marquardt algorithm, which brings the root mean square error (RMSE) of the reprojection error to ≤1.8 pixels.

[0032] The posture correction unit in the height calculation module calculates the torso direction vector. and vertical direction The included angle Achieving cosine compensation: The final height correction formula is: like The system prompts you to remain upright and pause the measurement.

[0033] The height calculation module also includes a ground normal vector estimation function: by estimating the ground normal vector from multiple three-dimensional points in the sole area. Perform plane fitting to solve the least squares problem: Obtain the optimal plane normal vector And adjust the height measurement direction from the Z-axis to along Direction, actual height calculated as follows: Suitable for applications where the ground tilt angle is ≤3°.

[0034] In the multi-view keypoint matching module, a basic matrix is ​​used. Establish epipolar constraints between the two viewpoints: for the left viewpoint point Point corresponding to the right view ,satisfy: in Let be the antisymmetric matrix of the translation vector. The RANSAC algorithm uses this constraint to select interior points and sets a reprojection error threshold. Pixels, maximum number of iterations This ensures robust matching.

[0035] The dynamic attitude stability assessment module defines key point motion stability indices. For the past Standard deviation of the location of a key point within a frame: When the key points of the hip or ankle If the attitude is deemed unstable, the system delays measurement until all key points have been measured for a continuous period of 0.5 seconds. .

[0036] Anomaly detection and fault tolerance mechanisms activate a mirror interpolation repair strategy when the single-view keypoint missing rate exceeds 40%: Let the left keypoint be... If missing, then use the right-side symmetrical point. The plane of the human body's central axis, formed by the tip of the nose, base of the neck, and mid-hip, has a normal vector. Calculate the mirror point: in This is the nearest projection point on the central plane. After repair, the reasonableness is verified by combining the prior bone length (e.g., shoulder width ≈ 0.25 × height). If the error exceeds ±15mm, it is marked as low confidence and manual review is prompted.

[0037] A non-contact height measurement method based on a multi-view image fusion non-contact automatic height measurement system includes the following steps: (a) The user guidance process is initiated by detecting the entry of a human body through infrared + ToF composite sensing; (b) Continuously monitor attitude stability indicators ,satisfy Synchronous shooting is triggered after lasting for more than 0.5 seconds; (c) Perform image distortion correction, using the Brown-Conrady model to correct radial and tangential distortions: in $, typical distortion coefficient ; (d) Use HRNet to extract key points, and SuperGlue+RANSAC to complete the matching; (e) The three-dimensional coordinates are reconstructed using a combination of EPNP and Bundling Adjustment. (f) Calculation compensate; (g) Activate the anomaly repair mechanism; (h) Collect 3 frames continuously, remove outliers with a deviation > 0.8 cm, and take the mean; (i) Perform privacy processing: The original image is Gaussian blurred (kernel function) Post-overwrite storage; (j) The results are uploaded in encrypted form, in compliance with GDPR and the Personal Information Protection Act.

[0038] The method involved testing 100 subjects aged 6–80 years and with a height of 100–200 cm under conditions of temperature 15–30℃, relative humidity 30%–70%, and light intensity 300–1000 lx. The systematic mean absolute error (MAE) was: Repeated measures standard deviation coefficient of variation The total time for a single process It also meets all the technical requirements for accuracy, stability and privacy security in the medical device standard YY / T 0654-2020.

[0039] The entire operation of this invention's system is fully automatic and unattended. From the user entering the measurement area to the final output of accurate height data, the entire process is completed within 3 seconds, requiring no human intervention. Its operation can be divided into the following seven stages, strictly executed according to a time sequence and logical control flow: Phase 1: User detection and entry recognition (t = 0~0.5 s) When the subject approaches the measuring device, an infrared proximity sensor or a ToF depth camera installed in front monitors spatial changes in real time. The system determines whether an object has entered the preset detection area (within 0.8 to 1.5 meters of the device). Once a human body is detected, the main control unit is immediately activated, and the vision subsystem in standby mode is started. Simultaneously, a voice prompt is triggered: "Welcome to the height measurement. Please stand in the center of the ground circle."

[0040] Phase 2: Attitude guidance and positioning calibration (t = 0.5–1.2 s) A ground laser projector projects a 60cm diameter ring of light to indicate the standard standing position; The ring-shaped LED status light illuminates yellow, indicating that positioning is in progress; The system initially acquires low-resolution images using dual cameras, and then analyzes the foot position in real time using ToF data. If both feet are within ±8cm of the aperture → Proceed to the next step; If the deviation is too large or one foot is lifted → please step both feet into the circle and continue to guide; Simultaneously, the dynamic posture stability assessment module is activated to continuously acquire video frames and calculate the displacement variance of key points in the hip and shoulder: when If the duration exceeds 0.5 seconds, the posture is considered stable.

[0041] Phase 3: Multi-view synchronous image acquisition (t = 1.2~1.4 s) Control signals are sent to two CMOS cameras (left and right) to trigger hardware-level synchronous shooting. The camera parameters are configured as follows: Resolution: 1920×1080 Frame rate: 30fps Exposure time: Automatic adjustment (ISO 100~800) Supplemental lighting: The LED array will be activated based on the ambient illuminance (target illuminance ≥ 300 lux). Image acquisition time difference ≤ 1ms, ensuring simultaneous imaging; The original image is transmitted to the edge computing unit (such as NVIDIA Jetson AGX Xavier) via the MIPI interface.

[0042] Phase 4: Image processing and human anatomy analysis (t = 1.4–1.9 s) (1) Image preprocessing Lens distortion correction using the Brown-Conrady model: in Perform white balance and contrast enhancement to improve image consistency.

[0043] (2) Human body segmentation The pre-trained Mask R-CNN model (ResNet-50-FPN) is used to perform semantic segmentation on the image; Output a binary mask to accurately extract human body contours, with an IoU ≥ 98.2%; Inference time ≤ 65ms (after TensorRT acceleration).

[0044] (3) Key point detection Seventeen COCO standard keypoints were detected using the HRNet-W48 network: nose tip, eye, shoulder, elbow, wrist, hip, knee, and ankle. Input dimensions: 256×192, output key point coordinates in heatmap format; Single frame processing time ≤ 80ms.

[0045] Phase 5: Cross-view matching and 3D reconstruction (t = 1.9–2.3 s) (1) Key point matching Extract key points and their surrounding features from the left and right images; Cross-view feature matching is performed using the SuperPoint + SuperGlue algorithm; Combining polar constraints Verify the correspondence; The RANSAC algorithm is used to remove false matches (threshold 2 pixels, 200 iterations) and retain interior points.

[0046] (2) Three-dimensional coordinate reconstruction For each pair of successfully matched 2D key points, triangulation is performed using camera calibration parameters; Camera projection model: The initial estimation uses the EPnP algorithm to solve for the attitude; Further Bundle Adjustment is performed to optimize reprojection error: This achieves a 3D reconstruction accuracy of ±2.5mm RMS.

[0047] (3) Anomaly repair mechanism (optional) If the missing keypoint rate at a certain viewpoint is >40%, enable mirror interpolation: The rationality was verified by combining the prior knowledge of bone length.

[0048] Phase 6: Height Calculation and Intelligent Compensation (t = 2.3–2.6 s) (1) Preliminary height calculation Take the head vertex in three-dimensional space (The value is obtained by interpolating upwards from the tip of the nose); Take the midpoint of the lowest point of the soles of both feet Calculate the Euclidean distance: (2) Attitude tilt compensation Construct the torso direction vector: Calculate the angle with the vertical direction : Correction formula: (3) Ground tilt compensation Find the normal vector of a plane fitted to multiple points on the sole of the foot. ; Final height is the projected length: Phase 7: Data fusion, privacy processing, and result output (t = 2.6–3.0 s) (1) Multi-frame fusion The system continuously collects 3 frames of valid data; Outliers with a bias > 0.8 cm were removed; Take the arithmetic mean of the remaining frames as the final result: (2) Privacy protection Gaussian blur is applied to the original image immediately. Or pixelated; Clear the memory cache and physically destroy the original image data; (3) Results display and uploading Height is displayed on the LCD screen; The LED light turns green, and the voice announces the height. Data is transmitted to the backend system via encrypted Wi-Fi (supports HTTPS + FHIR protocol). Supports printing barcode labels or synchronizing with electronic health records.

[0049] Python code for a non-contact automatic height measurement system based on multi-view image fusion: mport cv2 import numpy as np import math System parameters alpha = 2.0 Adaptive histogram equalization contrast enhancement limiting factor gamma = 1.0 (Height correction factor) theta_upright = 180 Standard value of joint angle in upright position delta_theta = 8 Allowable angular deviation H_std = 1.0 Standard reference height (meters) calibration_period = 15 Calibration period (days) Analog camera parameters k1, k2, p1, p2 = 0.1, 0.01, 0.001, 0.001 Camera distortion parameters K = np.array([[1000, 0, 960], [0, 1000, 540], [0, 0, 1]]) (Intrinsic parameter matrix) R = np.eye(3) Extrinsic parameter rotation matrix t = np.array([[0], [0], [0]]) Extrinsic parameter translation vector f = 1000 (Camera focal length) beta = math.radians(30) The angle between the camera's optical axis and the horizontal direction. hc = 1.5 Camera installation height Image preprocessing module def preprocess_image(image): Noise reduction processing def adaptive_median_filter(image): Simple Simulation Adaptive Median Filtering return cv2.medianBlur(image, 3) denoised_image = adaptive_median_filter(image) Image enhancement clahe = cv2.createCLAHE(clipLimit=alpha, tileGridSize=(8, 8)) enhanced_image = clahe.apply(denoised_image) Distortion correction height, width = image.shape[:2] corrected_image = np.zeros_like(image) for y in range(height): for x in range(width): r2 = x 2 + y 2 x_corrected = x (1 + k1 r2 + k2 r2 2) + 2 p1 x y + p2 (r2 + 2 x 2) y_corrected = y (1 + k1 r2 + k2 r2 2) + 2 p2 x y + p1 (r2 + 2 y 2) if 0<= x_corrected delta_theta: print("Please adjust your posture to an upright position") return False return True Cross-device data consistency calibration module def cross_device_calibration(): The system accuracy is automatically verified daily using a standard mannequin, triggering recalibration (simulated). print("Perform cross-device data consistency calibration") main function def main(): Simulated multi-view image acquisition image1 = cv2.imread('image1.jpg', cv2.IMREAD_GRAYSCALE) image2 = cv2.imread('image2.jpg', cv2.IMREAD_GRAYSCALE) Image preprocessing preprocessed_image1 = preprocess_image(image1) preprocessed_image2 = preprocess_image(image2) Multi-view image fusion fused_image = fuse_images(preprocessed_image1, preprocessed_image2) Attitude detection and correction if not detect_pose(fused_image): return 3D space reconstruction `image_points = np.array([[0, 0], [100, 100]])` This is a simple simulation of image points. object_points = reconstruct_3d(image_points) Height calculation height = calculate_height(fused_image, beta, f, hc, gamma) Data fusion and result output measurements = [height] average_height = sliding_average_filter(measurements) if average_height is not None: print(f"Measure height: {round(average_height)"); 100, 1)} cm") System calibration dpixel_std = 100 A simple simulation of the vertical pixel distance of a reference object. f_calibrated = calibrate_system(dpixel_std, beta, hc, H_std) print(f"Calibrated focal length: {f_calibrated}") Cross-device data consistency calibration cross_device_calibration() if __name__ == "__main__": main Image preprocessing module: Performs noise reduction, enhancement, and distortion correction on the acquired images.

[0050] Multi-view image fusion module: Features are extracted and matched using the ORB algorithm, and then image fusion is performed based on the Laplacian pyramid method.

[0051] 3D space reconstruction module: The 3D coordinates are initialized using the EPNP algorithm.

[0052] Height Calculation Module: Calculates and corrects a person's actual height using the principle of triangulation.

[0053] Data fusion and result output module: Performs moving average filtering on three consecutive valid measurement values ​​and outputs the results with an accuracy of 0.1cm.

[0054] System calibration module: Periodically calibrates the system and updates the camera's focal length parameters.

[0055] Posture detection and correction module: Detects human posture and issues a prompt if it does not meet the requirements.

[0056] Cross-device data consistency calibration module: Calibrates daily.

[0057] Test case To verify the measurement accuracy, repeatability, stability, and environmental adaptability of the non-contact automatic height measurement system based on multi-view image fusion of this invention, a series of control experiments were conducted. The detailed experimental design and results analysis are as follows.

[0058] I. Experimental Objective Verify the measurement accuracy of the system under standard conditions (compared with a standard medical electronic height meter). Evaluation system repeatability precision (coefficient of variation for multiple measurements by a single person); The robustness of the test system under different postures, lighting conditions, and non-ideal ground tilt conditions was tested. Verify the effectiveness of privacy protection mechanisms and data security; The long-term operational stability and the effectiveness of the automatic calibration function were examined.

[0059] II. Test Equipment and Environment

[0060] III. Test Subjects Human subjects: 100 in total, with the following age distribution: Children's group (6-12 years old): 30 people Adult group (18-60 years old): 50 people Senior group (65-80 years old): 20 people Height range: 105.3 cm ~ 198.7 cm, with a balanced gender ratio.

[0061] Standard doll: Used for automated testing and long-term stability verification.

[0062] All participants signed informed consent forms, and the trial process complied with ethical review requirements.

[0063] IV. Test Methods Experiment 1: Static accuracy test (standard standing posture) step: After removing their shoes, the subjects stood in the center of the positioning aperture of the system under test, with their feet together, looking straight ahead and remaining upright. The system automatically guides and acquires images and outputs height values. ; The same subject was immediately subjected to three manual measurements using an HC-800 electronic height meter, and the average value was taken as the reference value. ; Calculate the deviation for each case: .

[0064] index: Mean Absolute Error (MAE): Maximum Error Correlation coefficient result: Indicator values MAE 0.41 ± 0.12 cm Max Error 0.73 cm (occurred while an 8-year-old child was moving quickly) Pearson$r$0.9987 (p<0.001) The results show that the system's measured values ​​are highly consistent with those of the standard equipment, meeting the medical-grade accuracy requirements of YY / T 0654-2020.

[0065] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A non-contact automatic height measurement system based on multi-view image fusion, characterized in that, It includes a multi-view image acquisition module, equipped with at least two wide-angle CMOS cameras with a resolution of 1920×1080 and a frame rate of 30fps, installed on the left and right sides of the measurement area, with a horizontal angle of 45°±3°, a height of 1.5m±0.05m, a spacing of 1.2m, and a lens field of view ≥85° (H)×55° (V). It supports hardware-level synchronous triggering with a time deviation ≤1ms and is used to acquire frontal and side images of the subject. Image preprocessing module: Receives the raw images transmitted from the multi-view image acquisition module, performs denoising on the images using an adaptive median filtering algorithm, with the filter window size dynamically adjusted according to the noise density of local image regions; performs image enhancement using the adaptive histogram equalization (CLAHE) method, with a contrast enhancement limiting factor of [missing information]. The value range is 1.0-4.0; distortion correction is performed using pre-calibrated camera distortion parameters. The image pixel coordinates are determined according to the following formula. Perform correction: in, Multi-view image fusion module: This module uses a feature matching algorithm to extract and match features from preprocessed multi-view images, then fuses the matched images to produce a fused image containing complete human information. The fusion process employs a Laplacian pyramid-based fusion method, calculating the pixel values ​​of the fused image using the following formula. : in, These are the pixel values ​​of the two images to be merged. and The weights are calculated based on the image's sharpness and credibility. ; The 3D spatial reconstruction module is based on a pre-calibrated intrinsic parameter matrix. and external references ,use The algorithm initializes the 3D coordinates and achieves this by minimizing the reprojection error. optimization; Height Calculation Module: Combining pre-calibrated system parameters and extracted human feature point information, this module uses the principle of triangulation to calculate the actual height of a person. Assume the angle between the camera's optical axis and the horizontal direction is... The vertical pixel distance between the head vertex and the foot point on the image plane is Calculate a person's actual height using the following formula: At the same time, taking into account the physiological curvature of the human body, the calculation results are corrected, with a correction factor of [value missing]. The value range is 0.99-1.01, which is the corrected height. ; The data fusion and result output module performs a moving average filter on three consecutive valid measurement values ​​and finally outputs the result with an accuracy of 0.1cm. System calibration module: Performs periodic system calibration to ensure measurement accuracy; uses a standard height of... The system was calibrated using a reference object with a height accuracy of ±0.05 cm. During calibration, various system parameters were remeasured and updated, including the camera's focal length. Installation location and angle; calculate the calibrated focal length using the following formula. : in, It is the vertical pixel distance of the reference object on the image plane; the calibration cycle is once every half month or adjusted according to actual usage. The pose detection and correction module analyzes and fuses human pose information in images to determine whether the human body is in an upright position; it uses the OpenPose algorithm to detect the joint angles of the human body. When the system detects that the human posture does not meet the requirements for standing, it issues a voice prompt to guide the subject to adjust their posture until the correct measurement posture is achieved. The cross-device data consistency calibration module automatically verifies system accuracy and triggers recalibration daily using a standard mannequin. Data transmission module: Supports the transmission of measured height data to other devices or systems via wired or wireless networks, with a transmission rate of no less than 5Mbps, to achieve data sharing and remote monitoring; Display module: Uses an LCD screen to display the measured human height data, measurement time, and accuracy prompts for the measurement results. The screen size is no less than 10 inches, and the resolution is no less than 1280×800 pixels, to facilitate user viewing and operation.

2. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, In the three-dimensional space reconstruction module, any three-dimensional key point In the Projected coordinates on the imaging plane of each camera Satisfying the pinhole camera projection model: in As a scale factor, For the first The intrinsic parameter matrix of each camera.

3. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, The three-dimensional space reconstruction module further employs a nonlinear optimization method—beam adjustment. Construct an objective function to minimize the reprojection error across all viewpoints: in This represents the projection function with distortion correction. For visibility weights, if the first... Point at The value is 1 if it is visible from the viewpoint, and the optimization variables include the 3D coordinates of all key points. The camera pose parameters were solved using the Levenberg-Marquardt algorithm, which brought the root mean square error (RMSE) of the reprojection error to ≤1.8 pixels.

4. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, The posture correction unit in the height calculation module calculates the torso direction vector. and vertical direction The included angle Achieving cosine compensation: The final height correction formula is: like The system prompts you to remain upright and pause the measurement.

5. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, The height calculation module also includes a ground normal vector estimation function: by estimating multiple three-dimensional points in the sole area. Perform plane fitting to solve the least squares problem: Obtain the optimal plane normal vector And adjust the height measurement direction from the Z-axis to along Direction, actual height calculated as follows: Suitable for applications where the ground tilt angle is ≤3°.

6. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, The multi-view keypoint matching module uses a basic matrix. Establish epipolar constraints between the two viewpoints: for the left viewpoint point Point corresponding to the right view ,satisfy: in Let be the antisymmetric matrix of the translation vector; the RANSAC algorithm uses this constraint to select interior points and sets a reprojection error threshold. Pixels, maximum number of iterations This ensures robust matching.

7. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, The dynamic attitude stability assessment module defines key point motion stability indices. For the past Standard deviation of the location of a key point within a frame: When the key points of the hip or ankle If the attitude is deemed unstable, the system delays measurement until all key points have been measured for a continuous period of 0.5 seconds. .

8. The non-contact automatic height measurement system based on multi-view image fusion according to claim 1, characterized in that, The anomaly detection and fault tolerance mechanism activates a mirror interpolation repair strategy when the single-view keypoint missing rate exceeds 40%: Let the left keypoint be... If missing, then use the right-side symmetrical point. Normal vector to the central plane of the human body (formed by the tip of the nose, base of the neck, and mid-hip) Calculate the mirror point: in The nearest projection point on the central plane is used; after repair, the rationality is verified by combining the prior bone length. If the error exceeds ±15mm, it is marked as low confidence and manual review is prompted.

9. A non-contact height measurement method using the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: (a) The user guidance process is initiated by detecting the entry of a human body through infrared + ToF composite sensing; (b) Continuously monitor attitude stability indicators ,satisfy Synchronous shooting is triggered after lasting for more than 0.5 seconds; (c) Perform image distortion correction, using the Brown-Conrady model to correct radial and tangential distortions: in $, typical distortion coefficient ; (d) Use HRNet to extract key points, and SuperGlue+RANSAC to complete the matching; (e) The three-dimensional coordinates are reconstructed using a combination of EPNP and Bundling Adjustment. (f) Calculation compensate; (g) Activate the anomaly repair mechanism; (h) Collect 3 frames continuously, remove outliers with a deviation > 0.8 cm, and take the mean; (i) Perform privacy processing: The original image is Gaussian blurred (kernel function) Post-overwrite storage; (j) The result is uploaded in encrypted form.

10. The method according to claim 9, characterized in that, The method was tested on 100 subjects aged 6–80 years and with a height of 100–200 cm under conditions of temperature 15–30℃, relative humidity 30%–70%, and light intensity 300–1000 lx. The systematic mean absolute error (MAE) was: Repeated measures standard deviation coefficient of variation The total time for a single process It also meets all the technical requirements for accuracy, stability and privacy security in the medical device standard YY / T 0654-2020.

Images