Superficial ultrasonic scanning path planning method and device based on SVD-improved RRT algorithm

Abstract

The invention relates to the technical field of ultrasonic medical robots, and discloses a superficial ultrasonic scanning path planning method and device based on an SVD-improved RRT algorithm. The superficial ultrasonic scanning path planning method is applied to ultrasonic medical robot equipment, and specifically comprises the following steps: S101, receiving a data acquisition request sent by a main control end, controlling a Kinect camera to acquire a superficial depth image, processing 20 groups of calibration board images through a Matlab calibration tool box, and executing a corner detection algorithm to identify checkerboard corners, and optimizing and solving the internal reference matrix of the infrared camera. According to the method, SVD plane fitting and an improved RRT algorithm are fused, the precision and efficiency of an ultrasonic scanning path are remarkably improved, in the point cloud processing stage, statistical filtering and voxel gravity center down-sampling synergistically act, point cloud noise is reduced by 90%, meanwhile, more than 95% of three-dimensional features are reserved, a high-fidelity data basis is provided for path planning, and the method has good application prospects. SVD plane fitting forcibly passes through the center of mass, and the error is smaller than or equal to 0.5 mm

Description

Technical Field

[0001] This invention relates to the technical field of ultrasound medical robots, and more specifically, to a method and apparatus for shallow ultrasound scanning path planning based on the SVD-improved RRT algorithm. Background Technology

[0002] Ultrasound scanning technology, with its advantages of being radiation-free, providing real-time imaging, and being easy to operate, has become an important means of early cancer diagnosis and is widely used in clinical screening and interventional treatment. With the deep integration of medical robot technology and 3D imaging technology, ultrasound scanning is developing from traditional manual handheld probe operation to robot-assisted automation. Its core goal is to improve the consistency of scanning efficiency and diagnostic accuracy through standardized path planning. However, the complex three-dimensional morphology of human tissues, individual differences among patients, and possible tissue deformation during the scanning process have become key challenges to achieving high-precision automated scanning.

[0003] Existing technical solutions are mainly divided into three categories: remote control operation based on preset trajectories, autonomous path planning based on point clouds, and dynamic deformation compensation strategies. Remote control based on a preset trajectory achieves probe control through the collaboration of the doctor's console and the robotic arm. A typical example is the da Vinci system derivative solution from IntuitiveSurgical. This type of solution relies on the operator's experience to adjust the probe posture in real time. For example, Salcudean et al. first established a six-degree-of-freedom remote control system, and Lessard et al. expanded the workspace through a parallel structure robotic arm. However, it is still essentially dominated by human intervention and has the defects of low scanning efficiency and poor repeatability of results. Autonomous path planning schemes based on point clouds generate scanning trajectories through preoperative or intraoperative 3D modeling. For example, Huang et al. obtained point clouds on the skin surface from infrared cameras and used the center line of the region as the path. Kojcev et al. used RGB-D cameras to generate point clouds in real time and calculate the path. Tan et al. proposed an adaptive trajectory generation strategy based on 3D point cloud models. However, such schemes generally lack point cloud optimization processing and path smoothing algorithms. For example, Huang et al. did not consider point cloud denoising, and Kojcev et al. did not mention curve fitting, resulting in insufficient trajectory continuity and inability to adapt to shallow complex morphology. Dynamic deformation compensation strategies often employ image registration or finite element modeling techniques. The 3D ultrasound system developed by Hungr et al. tracks prostate movement through image registration, but a single registration takes 6-8 seconds, which cannot meet the real-time compensation requirements. Samei et al. achieved real-time ultrasound image registration based on a finite element model, but did not feed the deformation results back to the robot control. Tu et al. combined ultrasound and mixed reality technologies to calculate deformation through a shape prior model, but the dynamic compensation accuracy only reached the millimeter level, which still could not meet the sub-millimeter accuracy requirements of scanning. In addition, the probe trajectory tangent control method proposed by Welleweerd et al. did not combine point cloud normal vector calculation, and Wang et al.'s full-coverage scanning scheme lacked optimization of the point cloud fitting plane, all of which resulted in insufficient probe posture adjustment accuracy.

[0004] In summary, existing technologies have significant technical bottlenecks in areas such as the fine processing of 3D point cloud models, the geometric adaptability of path extraction algorithms, and the real-time performance of dynamic deformation compensation. There is an urgent need for a shallow ultrasonic scanning path planning method and device based on the SVD-improved RRT algorithm to solve the core problems of low scanning efficiency, high dependence on manual labor, and insufficient trajectory accuracy in traditional technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for shallow ultrasound scanning path planning based on SVD-improved RRT algorithm. By integrating SVD plane fitting and improved RRT algorithm, the accuracy and efficiency of ultrasound scanning path are significantly improved. In the point cloud processing stage, statistical filtering and voxel centroid downsampling work together to reduce point cloud noise by 90% while retaining more than 95% of the three-dimensional features, providing a high-fidelity data foundation for path planning, and aiming to solve the problems in the prior art.

[0006] This invention is implemented as follows: a shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm is applied to ultrasound medical robot equipment, specifically including the following steps: S101: Receives data acquisition requests from the main control unit, controls the Kinect camera to acquire shallow depth images, processes 20 sets of calibration board images using the Matlab calibration toolbox, executes a corner detection algorithm to identify checkerboard corners, and optimizes the solution of the infrared camera intrinsic parameter matrix, which includes the focal length. and optical center coordinates ; S102: Obtain the infrared depth value d and scaling factor through the infrared camera intrinsic parameter matrix. An initial point cloud is generated by calculating the coordinates of spatial points using a three-dimensional coordinate formula. ; Where d is the pixel depth value. The scaling factor for the depth value. , For the focal length of the infrared camera, , Using optical center coordinates, a statistical filtering algorithm is used to remove outliers from the initial point cloud to generate a fine point cloud. Then, a fast uniform downsampling algorithm is used to divide the space of the fine point cloud into a voxel grid, retaining the centroid of each grid. S103: Calculate the centroid of the fine point cloud and construct the covariance matrix using the singular value decomposition (SVD) algorithm. Take the eigenvector corresponding to the smallest eigenvalue of the covariance matrix as the plane normal vector. Extract the intersection line between the plane and the fine point cloud contour to form a closed circle. Divide the closed circle into 16 path points and calculate the normal vector of each path point through local surface fitting. S104: Based on the distance between the current tree node and the target point using the normal vectors of each path point. Dynamically adjust the sampling probability of the target point: ; in To control the parameters of the rate of change of the sampling probability to generate new nodes, when the tree approaches the target point, Decrease Increasing the sampling frequency of the target point accelerates the expansion of the tree towards the target region. As the tree moves further away from the target point... Increase The sampling ratio of random points is increased by reducing the cost of the new node and selecting a set of neighboring nodes within a preset radius. The node with the minimum cost is selected as the parent node by the cost function. A fifth-order B-spline curve is used to smooth the path of the parent node to generate path points. An ultrasound scanning path is generated through each path point. S105: Obtain the end-effector pose of the ultrasound medical robot and the camera observation pose, drive the calibration board to perform hand-eye calibration, input the ultrasound scanning path into the ultrasound medical robot, constrain the start / end position of the path and align the endpoint tangent vector with the point cloud normal, and execute the scanning trajectory to complete the shallow ultrasound scanning path planning.

[0007] Specifically, superficial areas generally include the breast, carotid artery, thyroid gland, etc. This technical solution enables the robot to complete ultrasound scanning path planning for superficial areas based on the SVD-improved RRT algorithm.

[0008] Furthermore, in S101, receiving a data acquisition request sent by the master control terminal includes: The system receives a connection request sent by the main control terminal through a preset network. The connection request is used to request the establishment of a connection with the electronic device. Detect whether the current account on the main control terminal is the target account; If the current account on the main control terminal is the target account, a connection is established with the main control terminal according to the connection request. After the connection is established, a data collection request sent by the main control terminal is received.

[0009] Furthermore, 20 sets of calibration board images were processed using the Matlab calibration toolbox, a corner detection algorithm was executed to identify checkerboard corners, and the intrinsic parameter matrix of the infrared camera was optimized, including: Using the Matlab camera calibration toolbox, 20 sets of pre-acquired chessboard calibration board images from different orientations and distances were systematically processed. For each set of images, a high-precision corner detection algorithm was executed to automatically identify and accurately lock the coordinates of all interior corner points at the intersection of black and white squares in the chessboard. By identifying all image corner points and their corresponding known 3D world coordinates, a complex camera imaging geometric model is constructed. Through nonlinear optimization algorithms, the core intrinsic parameter matrix of the infrared camera and the key lens distortion parameters are jointly optimized and solved.

[0010] Furthermore, in S102, a statistical filtering algorithm is used to remove outliers from the initial point cloud to generate a fine point cloud. Then, a fast uniform downsampling algorithm is used to divide the space of the fine point cloud into a voxel grid, including: For each point in the initial point cloud, calculate the average distance to its K neighboring points and the standard deviation of the distance distribution. Set a distance threshold as the average value plus three times the standard deviation. If the average distance between a point and its neighboring points exceeds this threshold, it is identified as a noise point and removed. To efficiently compress the initial point cloud size while preserving the key 3D structure, fast and uniform sampling based on voxel mesh is implemented. The appropriate voxel cube size is calculated according to the desired sampling density or the target point cloud reduction ratio. The entire point cloud space is divided into a regular 3D mesh of this size. Then, the geometric centroid of all points inside each non-empty voxel is calculated, and the centroid is output as the unique representative point of the voxel.

[0011] Furthermore, in S103, the eigenvector corresponding to the smallest eigenvalue of the covariance matrix is ​​taken as the plane normal vector through SVD decomposition. The intersection line of this plane and the fine point cloud contour is extracted to form a closed circle. The closed circle is divided into 16 path points, including: The eigenvector corresponding to the smallest eigenvalue of the matrix is ​​obtained by Singular Value Decomposition (SVD) and used as the plane normal vector. The fitting plane is forced to pass through the centroid, ensuring that the sum of the projected distances of all points to the plane is minimized. The fitting error is strictly controlled within 0.5mm, thus establishing a high-precision reference plane for path extraction. The contour circle is divided into 16 segments using the equal angle segmentation method. A segmentation point is taken as a path point every 22.5°. For each path point, the local surface normal vector is solved in its 5mm radius neighborhood using a quadratic surface fitting algorithm, which is used for real-time adjustment of the probe attitude.

[0012] Furthermore, in S104, a set of neighboring nodes within a preset radius is selected from the new node, and the node with the minimum cost is selected as the parent node through a cost function, including: After a new node is generated, all neighboring nodes within a set radius are filtered out, rather than just selecting the nearest node, using a formula. Calculate the path cost from each neighboring node to the new node, and select the node with the lowest cost as the parent node of the new node; Check if the path cost of neighboring nodes can be reduced by the new node. If so... Then update its parent node to achieve path reconnection and reduce overall cost.

[0013] Furthermore, in S101, a quintic B-spline curve is used to smooth the parent node path to generate path points. An ultrasound scanning path is then generated from these path points, including: The path smoothing stage uses fifth-order B-spline interpolation, with the following formula: ,in To control the vertices; Fifth-order B-spline basis functions This ensures the continuity of the first and second derivatives of the path, and that local adjustments only affect three adjacent curve segments. It adapts the path point distribution characteristics to non-uniform node vectors, balancing smoothness and accuracy. Implement endpoint constraints, including: Position constraints ensure that the starting and ending points of the path strictly correspond to the preset scanning start and target points, with an error ≤0.2mm; The attitude constraint is achieved by solving inverse kinematics and checking the effectiveness of joint space, so that the endpoint tangent vector is aligned with the point cloud normal vector, ensuring that the robot arm end posture fits the target surface. In addition, non-adjacent nodes are randomly selected, and if there is no collision when connecting straight lines, the intermediate nodes are deleted, thereby shortening the path length while satisfying the constraints.

[0014] Further, in S105, the end-effector pose of the ultrasound medical robot and the camera observation pose are obtained, and the calibration board is driven to perform hand-eye calibration, including: By solving the transformation relationship through matrix operations, the final hand-eye matrix calibration error is less than 0.8mm, achieving coordinate unification between the camera and the robot. The hand-eye transformation matrix B is solved based on the homogeneous transformation equation. ; If the average error between the planned path and the actual path in the x, y, and z directions based on the transformation matrix B is within 0.66 mm, 0.68 mm, and 0.82 mm respectively, then the shallow ultrasound scan path planning is completed by executing the scan trajectory.

[0015] Compared with the prior art, the shallow ultrasound scanning path planning method and device based on the SVD-improved RRT algorithm provided by the present invention has the following beneficial effects: 1. By integrating SVD plane fitting and the improved RRT algorithm, the accuracy and efficiency of ultrasound scanning path are significantly improved. In the point cloud processing stage, statistical filtering and voxel centroid downsampling work together to reduce point cloud noise by 90% while retaining more than 95% of the three-dimensional features, providing a high-fidelity data foundation for path planning. SVD plane fitting is forced to pass through the centroid with an error ≤0.5mm. Combined with the 16-part contour circle strategy, it ensures that the path points uniformly cover the key areas of the shallow surface. The improved RRT algorithm dynamically balances exploration and convergence efficiency through an adaptive target sampling probability function, which improves the convergence speed compared with the traditional RRT. The neighborhood reconnection mechanism combined with fifth-order B-spline smoothing shortens the path length and reduces curvature abrupt changes by 70%, greatly reducing the jitter of the robotic arm. The dual constraint of the endpoints ensures that the probe is accurately attached to the shallow surface and avoids scanning blind spots. 2. At the device level, the system aligns with clinical operational needs, achieving deep collaboration between hardware and software. The modular design of point cloud processing, path planning, RRT optimization, and control verification supports efficient execution of the algorithm chain. The processor analyzes depth images in real time via the Kinect interface and drives the UR5e robotic arm to execute the trajectory using EtherCAT communication. Hand-eye calibration error <0.8mm and three-axis trajectory error closed-loop verification constitute a clinical-grade accuracy assurance system, meeting the stringent requirements of medical standards for scanning accuracy. Path shortening strategies and real-time collision detection reduce invalid movements and improve single-scan efficiency. The quick-release probe structure adapts to sterile environments, while the error heatmap generation function provides operators with a quantitative assessment tool. The overall solution compresses planning time to within 5 minutes, significantly improving efficiency compared to manual planning and providing reliable technical support for automated ultrasound screening.

[0016] A shallow ultrasound scanning path planning device based on the SVD-improved RRT algorithm includes a storage device and a processor. The storage device stores a computer program, and the processor runs the computer program to enable the shallow ultrasound scanning path planning device to perform the aforementioned shallow ultrasound scanning path planning method.

[0017] Specifically, the storage device includes: Point cloud processing module: performs point cloud data acquisition, 3D coordinate calculation, statistical filtering, and voxel downsampling operations; Path planning module: Performs SVD plane fitting, closed contour segmentation, and path point normal vector calculation; RRT optimization module: performs adaptive target point sampling, neighborhood reconnection and fifth-order B-spline smoothing algorithm; Control verification module: performs endpoint position / attitude constraint verification and path shortening strategy; A processor is configured to run the computer program to achieve: The Kinect camera interface is used to acquire shallow depth images in real time, which in turn drive the point cloud processing module to generate optimized point clouds. The path planning module is invoked to extract 16 equally divided path points and their normal vectors to construct the initial scan path; The RRT optimization module is activated to generate a smooth robotic arm trajectory, and the endpoint inverse kinematics solution parameters are loaded. The trajectory is output to the UR5e robotic arm through the robot control interface, and the control verification is triggered simultaneously to monitor the pose error and collision detection in real time. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm proposed in this invention. Figure 2 The flowchart of the shallow ultrasound scanning path planning method based on SVD-improved RRT algorithm proposed in this invention is as follows: the statistical filtering algorithm is used to remove outliers in the initial point cloud to generate a fine point cloud, and then the space of the fine point cloud is divided into a voxel grid by a fast uniform downsampling algorithm. Figure 3 This is a schematic diagram of the shallow ultrasound scanning path planning device based on the SVD-improved RRT algorithm proposed in this invention. Figure 4 This is the operational logic diagram of the shallow ultrasound scanning path planning device based on the SVD-improved RRT algorithm proposed in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] The implementation of the present invention will be described in detail below with reference to specific embodiments.

[0021] In the accompanying drawings of this embodiment, the same or similar reference numerals correspond to the same or similar components. In the description of this invention, it should be understood that if terms such as "upper," "lower," "left," and "right" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting this invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0022] Reference Figure 1-2 As shown, a shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm is applied to ultrasound medical robot equipment, specifically including the following steps: S101: Receives data acquisition requests from the main control unit, controls the Kinect camera to acquire shallow depth images, processes 20 sets of calibration board images using the Matlab calibration toolbox, executes a corner detection algorithm to identify checkerboard corners, and optimizes the solution of the infrared camera intrinsic parameter matrix, which includes the focal length. and optical center coordinates After calibration, the reprojection errors of the RGB camera and the infrared camera were controlled within 0.42 pixels and 0.33 pixels respectively, ensuring the accuracy of subsequent 3D coordinate transformation. This included processing 20 sets of calibration board images using the Matlab calibration toolbox, executing a corner detection algorithm to identify checkerboard corners, and optimizing the solution of the infrared camera intrinsic parameter matrix, including: Using the Matlab camera calibration toolbox, 20 sets of pre-acquired chessboard calibration board images from different orientations and distances were systematically processed. For each set of images, a high-precision corner detection algorithm was executed to automatically identify and accurately lock the coordinates of all interior corner points at the intersection of black and white squares in the chessboard. All image corner points and their corresponding known 3D world coordinates were identified, and a complex camera imaging geometric model was constructed. Through nonlinear optimization algorithms, the core intrinsic parameter matrix of the infrared camera and the key lens distortion parameters were jointly optimized and solved. The radial distortion parameters were optimized to [2.4304, -39.0622] by the least squares method, ensuring that the transformation error from the world coordinate system to the camera coordinate system is <0.5 pixels. S102: Obtain the infrared depth value d and scaling factor through the infrared camera intrinsic parameter matrix. An initial point cloud is generated by calculating the coordinates of spatial points using a three-dimensional coordinate formula. ; Where d is the pixel depth value. The scaling factor for the depth value. , For the focal length of the infrared camera, , Using optical center coordinates, a statistical filtering algorithm is used to remove outliers from the initial point cloud to generate a fine point cloud. Then, a fast uniform downsampling algorithm is used to divide the space of the fine point cloud into a voxel grid, retaining the centroid of each grid. The process involves using a statistical filtering algorithm to remove outliers from the initial point cloud to generate a refined point cloud, and then using a fast uniform downsampling algorithm to divide the space of the refined point cloud into a voxel grid, including: For each point in the initial point cloud, calculate the average distance to its K neighboring points and the standard deviation of the distance distribution. Set a distance threshold as the average value plus three times the standard deviation. If the average distance between a point and its neighboring points exceeds this threshold, it is identified as a noise point and removed. To efficiently compress the initial point cloud size while preserving the key 3D structure, fast and uniform sampling based on voxel mesh is implemented. The appropriate voxel cube size is calculated according to the desired sampling density or the target point cloud reduction ratio. The entire point cloud space is divided into a regular 3D mesh of this size. Then, the geometric centroid of all points inside each non-empty voxel is calculated, and the centroid is output as the unique representative point of the voxel. S103: Calculate the centroid of the fine point cloud and construct the covariance matrix using the singular value decomposition (SVD) algorithm. Take the eigenvector corresponding to the smallest eigenvalue of the covariance matrix as the plane normal vector. Extract the intersection line between the plane and the fine point cloud contour to form a closed circle. Divide the closed circle into 16 path points and calculate the normal vector of each path point through local surface fitting. Specifically, the eigenvector corresponding to the smallest eigenvalue of the covariance matrix is ​​taken as the plane normal vector through SVD decomposition. The intersection line of this plane and the fine point cloud contour is extracted to form a closed circle. The closed circle is then divided into 16 path points, including: The eigenvector corresponding to the smallest eigenvalue of the matrix is ​​obtained by Singular Value Decomposition (SVD) and used as the plane normal vector. The fitting plane is forced to pass through the centroid, ensuring that the sum of the projected distances of all points to the plane is minimized. The fitting error is strictly controlled within 0.5mm, thus establishing a high-precision reference plane for path extraction. The contour circle is divided into 16 segments using the equal angle segmentation method. A segmentation point is taken at every 22.5° as a path point. For each path point, the local surface normal vector is solved in its 5mm radius neighborhood using a quadratic surface fitting algorithm, which is used for real-time adjustment of the probe attitude. S104: Based on the distance between the current tree node and the target point using the normal vectors of each path point. Dynamically adjust the sampling probability of the target point: ; in To control the parameters of the rate of change of the sampling probability to generate new nodes, when the tree approaches the target point, Decrease Increasing the sampling frequency of the target point accelerates the expansion of the tree towards the target region. As the tree moves further away from the target point... Increase The sampling ratio of random points is increased by reducing the cost of the new node and selecting a set of neighboring nodes within a preset radius. The node with the minimum cost is selected as the parent node by the cost function. A fifth-order B-spline curve is used to smooth the path of the parent node to generate path points. An ultrasound scanning path is generated through each path point. Among these steps, the neighboring node set within a preset radius is selected in the new node, and the node with the minimum cost is selected as the parent node through a cost function, including: After a new node is generated, all neighboring nodes within a set radius are filtered out, rather than just selecting the nearest node, using a formula. Calculate the path cost from each neighboring node to the new node, and select the node with the lowest cost as the parent node of the new node; Check if the path cost of neighboring nodes can be reduced by the new node. If so... Then update its parent node to achieve path reconnection and reduce overall cost; S105: Obtains the end-effector pose and camera observation pose of the ultrasound medical robot, drives the calibration board for hand-eye calibration, inputs the ultrasound scanning path into the ultrasound medical robot, constrains the start / end point positions of the path and aligns the endpoint tangent vectors with the point cloud normals, and executes the scanning trajectory to complete the shallow ultrasound scanning path planning. By integrating SVD plane fitting and an improved RRT algorithm, the accuracy and efficiency of the ultrasound scanning path are significantly improved. In the point cloud processing stage, statistical filtering and voxel centroid downsampling work together to reduce point cloud noise by 90% while retaining more than 95% of the 3D representation. Features provide a high-fidelity data foundation for path planning. SVD plane fitting forces the probe to pass through the centroid with an error ≤0.5mm. Combined with a 16-part contour circle strategy, it ensures that path points uniformly cover key areas of the shallow surface. The improved RRT algorithm dynamically balances exploration and convergence efficiency through an adaptive target sampling probability function, resulting in faster convergence than traditional RRT. The neighborhood reconnection mechanism combined with fifth-order B-spline smoothing shortens the path length and reduces curvature abrupt changes by 70%, significantly reducing robotic arm motion jitter. Dual endpoint constraints ensure precise contact between the probe and the shallow surface, avoiding scanning blind spots.

[0023] In this embodiment, in S101, receiving a data acquisition request sent by the master control terminal includes: The system receives connection requests sent by the master control unit through a preset network. These connection requests are used to request the establishment of a connection with the electronic device. Detect whether the current account on the main control terminal is the target account; If the current account on the master control terminal is the target account, it will connect to the master control terminal according to the connection request. After the connection is completed, it will receive the data collection request sent by the master control terminal.

[0024] In S104 of this embodiment, a quintic B-spline curve is used to smooth the parent node path to generate path points, and an ultrasound scanning path is generated through each path point, including: The path smoothing stage uses fifth-order B-spline interpolation, with the following formula: ,in To control the vertices; Fifth-order B-spline basis functions This ensures the continuity of the first and second derivatives of the path, and that local adjustments only affect three adjacent curve segments. It adapts the path point distribution characteristics to non-uniform node vectors, balancing smoothness and accuracy. Implement endpoint constraints, including: Position constraints ensure that the starting and ending points of the path strictly correspond to the preset scanning start and target points, with an error ≤0.2mm; The attitude constraint is achieved by solving inverse kinematics and checking the effectiveness of joint space, so that the endpoint tangent vector is aligned with the point cloud normal vector, ensuring that the robot arm end posture fits the target surface. In addition, non-adjacent nodes are randomly selected, and if there is no collision when connecting straight lines, the intermediate nodes are deleted, thereby shortening the path length while satisfying the constraints.

[0025] In S105 of this embodiment, obtaining the end-effector pose and camera observation pose of the ultrasound medical robot device, and driving the calibration board to perform hand-eye calibration, includes: By solving the transformation relationship through matrix operations, the final hand-eye matrix calibration error is less than 0.8mm, achieving coordinate unification between the camera and the robot. The hand-eye transformation matrix B is solved based on the homogeneous transformation equation. ; If the average error between the measured planned path and the actual path in the x, y, and z directions based on the transformation matrix B is within 0.66 mm, 0.68 mm, and 0.82 mm respectively, then the shallow ultrasound scan path planning is completed by executing the scan trajectory.

[0026] Specifically, this technical solution's path shortening strategy and real-time collision detection reduce unnecessary movement and improve single-scan efficiency. The quick-release probe structure is suitable for sterile environments, while the error heatmap generation function provides operators with a quantitative assessment tool. The overall solution reduces planning time to within 5 minutes, significantly improving efficiency compared to manual planning and providing reliable technical support for automated ultrasound screening. Compared with traditional methods, the error of 3D point cloud reconstruction is reduced by 40%, and the average error of path planning is controlled within 0.8mm.

[0027] Efficiency optimization: Automatic path generation time ≤200ms, 5 times more efficient than manual planning, suitable for clinical batch scanning.

[0028] Enhanced adaptability: Through point cloud model optimization and dynamic fitting, it can adapt to the superficial morphology of patients with different body types, with a scanning coverage of 99.2%.

[0029] Cost reduction: Only a Kinect camera and a general industrial robot are required, eliminating the need for expensive equipment such as laser scanning, reducing system costs by 60% compared to similar solutions.

[0030] This technical solution significantly improves the accuracy and efficiency of ultrasound scanning paths by integrating SVD plane fitting and an improved RRT algorithm. In the point cloud processing stage, statistical filtering and voxel centroid downsampling work together to reduce point cloud noise by 90% while retaining more than 95% of the three-dimensional features, providing a high-fidelity data foundation for path planning. The processor analyzes the depth image in real time through the Kinect interface and drives the UR5e robotic arm to execute the trajectory via EtherCAT communication. The hand-eye calibration error of <0.8mm and the three-axis trajectory error closed-loop verification constitute a clinical-grade accuracy assurance system, meeting the stringent requirements of medical standards for scanning accuracy. The path shortening strategy and real-time collision detection reduce invalid movements and improve the efficiency of a single scan. The quick-release probe structure is adapted to the sterile environment, and the error heatmap generation function provides operators with a quantitative evaluation tool. The overall solution compresses the planning time to within 5 minutes, which is more efficient than manual planning and provides reliable technical support for automated ultrasound screening.

[0031] Reference Figure 3-4 As shown, a shallow ultrasound scanning path planning device based on the SVD-improved RRT algorithm includes a storage device and a processor. The storage device stores a computer program, and the processor runs the computer program to enable the shallow ultrasound scanning path planning device to perform the aforementioned shallow ultrasound scanning path planning method.

[0032] In this embodiment, the storage device includes: Point cloud processing module: performs point cloud data acquisition, 3D coordinate calculation, statistical filtering, and voxel downsampling operations; Path planning module: Performs SVD plane fitting, closed contour segmentation, and path point normal vector calculation; RRT optimization module: performs adaptive target point sampling, neighborhood reconnection and fifth-order B-spline smoothing algorithm; Control verification module: performs endpoint position / attitude constraint verification and path shortening strategy; processor, configured to run the computer program to implement: The Kinect camera interface is used to acquire shallow depth images in real time, which in turn drive the point cloud processing module to generate optimized point clouds. The path planning module is invoked to extract 16 equally divided path points and their normal vectors to construct the initial scan path; The RRT optimization module is activated to generate a smooth robotic arm trajectory, and the endpoint inverse kinematics solution parameters are loaded. The trajectory is output to the UR5e robotic arm through the robot control interface, and the control verification is triggered simultaneously to monitor the pose error and collision detection in real time.

[0033] This device is designed to meet clinical operational needs, enabling deep collaboration between hardware and software. The modular design of point cloud processing, path planning, RRT optimization, and control verification supports efficient execution of the algorithm chain. The processor analyzes depth images in real time through the Kinect interface and drives the UR5e robotic arm to execute the trajectory via EtherCAT communication. The hand-eye calibration error of <0.8mm and the closed-loop verification of the three-axis trajectory error constitute a clinical-grade accuracy assurance system, meeting the stringent requirements of medical standards for scanning accuracy.

[0034] In terms of hardware implementation, the sensor system constructed in this invention uses a Kinectv2 camera for shallow depth image acquisition. This camera has a depth resolution of 512×424 and a frame rate of 30fps. It connects to an industrial computer equipped with an Intel i7-8700 processor and 16GB of memory via a USB 3.0 interface for efficient data transmission and processing. The robot system uses a UR5e six-axis robot with a repeatability accuracy of ±0.1mm. Its end effector is equipped with a customized ultrasound probe holder, providing 360° attitude adjustment capability to flexibly adapt to scanning needs. The ultrasound equipment uses a Mindray DC-70S ultrasound machine with a 7.5-10MHz linear array probe for real-time acquisition and verification of ultrasound images. This hardware configuration supports real-time computation of the improved RRT algorithm in areas such as neighborhood search and path reconnection, meeting the timeliness requirements of clinical scanning.

[0035] In the software implementation phase, point cloud processing is performed using the PCL library. Statistical filtering algorithms are employed to remove random and Gaussian noise, and a fast uniform downsampling algorithm is used to reduce the number of points in the cloud. While preserving 3D structural features, the point cloud reconstruction time is controlled to within 150ms. Path planning is based on the MATLAB platform. First, the scanning start point and target point are determined. An improved RRT algorithm is used for path planning. An adaptive target point sampling strategy dynamically adjusts the target point sampling probability. Path cost is optimized by combining neighborhood search and path reconnection. Path shortening is then performed, and fifth-order B-spline interpolation is used to achieve path smoothing, ensuring path continuity and smoothness. The path generation cycle is controlled to within 80ms. Robot control relies on the ROS system, sending trajectory points at a 100Hz sampling frequency. Combined with inverse kinematics solving and joint space validity checks, the control latency is maintained below 30ms, ensuring accurate execution of scanning actions.

[0036] The actual workflow follows a rigorous logical progression. First, the Kinect camera acquires shallow depth images, and a 3D point cloud model is reconstructed using pre-calibrated parameters. Next, denoising and downsampling optimization are performed on the original point cloud. Then, an improved RRT algorithm is used to generate an initial path, which is optimized, shortened, and smoothed to obtain the final scanning trajectory. Afterward, the path coordinates are transformed to the robot's base coordinate system using a hand-eye calibration matrix, driving the UR5e robot to perform the scanning operation. Finally, ultrasound images are acquired in real time, and the planned path is compared with the actual trajectory to continuously verify errors and optimize relevant parameters, forming a complete and closed-loop shallow ultrasound scanning path planning and execution system. This workflow, leveraging the advantages of the improved RRT algorithm, effectively improves the efficiency and quality of path planning in complex environments, meeting the dual requirements of efficiency and accuracy in clinical applications.

[0037] In this embodiment, the entire operation process can be controlled by a computer to provide signal feedback and implement the steps sequentially. These are all conventional knowledge in current automation control, and will not be elaborated on in this embodiment.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A shallow ultrasound scanning path planning method based on SVD-improved RRT algorithm, characterized in that, Applied to ultrasound medical robot equipment, the specific steps include: S101: Receives data acquisition requests from the main control unit, controls the Kinect camera to acquire shallow depth images, processes 20 sets of calibration board images using the Matlab calibration toolbox, executes a corner detection algorithm to identify checkerboard corners, and optimizes the solution of the infrared camera intrinsic parameter matrix, which includes the focal length. and optical center coordinates ; S102: Obtain the infrared depth value d and scaling factor through the infrared camera intrinsic parameter matrix. An initial point cloud is generated by calculating the coordinates of spatial points using a three-dimensional coordinate formula. ； Where d is the pixel depth value. The scaling factor for the depth value. , For the focal length of the infrared camera, , Using optical center coordinates, a statistical filtering algorithm is used to remove outliers from the initial point cloud to generate a fine point cloud. Then, a fast uniform downsampling algorithm is used to divide the space of the fine point cloud into a voxel grid, retaining the centroid of each grid. S103: Calculate the centroid of the fine point cloud and construct the covariance matrix using the singular value decomposition (SVD) algorithm. Take the eigenvector corresponding to the smallest eigenvalue of the covariance matrix as the plane normal vector. Extract the intersection line between the plane and the fine point cloud contour to form a closed circle. Divide the closed circle into 16 path points and calculate the normal vector of each path point through local surface fitting. S104: Based on the distance between the current tree node and the target point using the normal vectors of each path point. Dynamically adjust the sampling probability of the target point: ； in To control the parameters of the rate of change of the sampling probability to generate new nodes, when the tree approaches the target point, Decrease Increasing the sampling frequency of the target point accelerates the expansion of the tree towards the target region. As the tree moves further away from the target point... Increase The sampling ratio of random points is increased by reducing the cost of the new node and selecting a set of neighboring nodes within a preset radius. The node with the minimum cost is selected as the parent node by the cost function. A fifth-order B-spline curve is used to smooth the path of the parent node to generate path points. An ultrasound scanning path is generated through each path point. S105: Obtain the end-effector pose of the ultrasound medical robot and the camera observation pose, drive the calibration board to perform hand-eye calibration, input the ultrasound scanning path into the ultrasound medical robot, constrain the start / end position of the path and align the endpoint tangent vector with the point cloud normal, and execute the scanning trajectory to complete the shallow ultrasound scanning path planning.

2. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 1, characterized in that, In S101, the data acquisition request sent by the master control terminal is received, including: The system receives a connection request sent by the main control unit through a preset network. The connection request is used to request the establishment of a connection with the electronic device. Detect whether the current account on the main control terminal is the target account; If the current account on the main control terminal is the target account, a connection is established with the main control terminal according to the connection request. After the connection is established, a data collection request sent by the main control terminal is received.

3. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 2, characterized in that, Twenty sets of calibration board images were processed using the Matlab calibration toolbox. A corner detection algorithm was executed to identify checkerboard corners, and the intrinsic parameter matrix of the infrared camera was optimized. This included: Using the Matlab camera calibration toolbox, 20 sets of pre-acquired chessboard calibration board images from different orientations and distances were systematically processed. For each set of images, a high-precision corner detection algorithm was executed to automatically identify and accurately lock the coordinates of all interior corner points at the intersection of black and white squares in the chessboard. By identifying all image corner points and their corresponding known 3D world coordinates, a complex camera imaging geometric model is constructed. Through nonlinear optimization algorithms, the core intrinsic parameter matrix of the infrared camera and the key lens distortion parameters are jointly optimized and solved.

4. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 3, characterized in that, In S102, a statistical filtering algorithm is used to remove outliers from the initial point cloud to generate a refined point cloud. Then, a fast uniform downsampling algorithm is used to divide the space of the refined point cloud into a voxel grid, including: For each point in the initial point cloud, calculate the average distance to its K neighboring points and the standard deviation of the distance distribution. Set a distance threshold as the average value plus three times the standard deviation. If the average distance between a point and its neighboring points exceeds this threshold, it is identified as a noise point and removed. To efficiently compress the initial point cloud size while preserving the key 3D structure, fast and uniform sampling based on voxel mesh is implemented. The appropriate voxel cube size is calculated according to the desired sampling density or the target point cloud reduction ratio. The entire point cloud space is divided into a regular 3D mesh of this size. Then, the geometric centroid of all points inside each non-empty voxel is calculated, and the centroid is output as the unique representative point of the voxel.

5. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 4, characterized in that, In S103, the eigenvector corresponding to the smallest eigenvalue of the covariance matrix is ​​taken as the plane normal vector after SVD decomposition. The intersection line of this plane and the fine point cloud contour is extracted to form a closed circle. The closed circle is divided into 16 path points, including: The eigenvector corresponding to the smallest eigenvalue of the matrix is ​​obtained by Singular Value Decomposition (SVD) and used as the plane normal vector. The fitting plane is forced to pass through the centroid, ensuring that the sum of the projected distances of all points to the plane is minimized. The fitting error is strictly controlled within 0.5mm, thus establishing a high-precision reference plane for path extraction. The contour circle is divided into 16 segments using the equal angle segmentation method. A segmentation point is taken as a path point every 22.5°. For each path point, the local surface normal vector is solved in its 5mm radius neighborhood using a quadratic surface fitting algorithm, which is used for real-time adjustment of the probe attitude.

6. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 5, characterized in that, In S104, a set of neighboring nodes within a preset radius is selected from the new node, and the node with the minimum cost is selected as the parent node using a cost function, including: After a new node is generated, all neighboring nodes within a set radius are filtered out, rather than just selecting the nearest node, using a formula. Calculate the path cost from each neighboring node to the new node, and select the node with the lowest cost as the parent node of the new node; Check if the path cost of neighboring nodes can be reduced by the new node. If so... Then update its parent node to achieve path reconnection and reduce overall cost.

7. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 6, characterized in that, In S101, a quintic B-spline curve is used to smooth the parent node path to generate path points. The ultrasound scanning path is then generated from these path points, including: The path smoothing stage uses fifth-order B-spline interpolation, with the following formula: ,in To control the vertices; Fifth-order B-spline basis functions This ensures the continuity of the first and second derivatives of the path, and that local adjustments only affect three adjacent curve segments. It adapts the path point distribution characteristics to non-uniform node vectors, balancing smoothness and accuracy. Implement endpoint constraints, including: Position constraints ensure that the starting and ending points of the path strictly correspond to the preset scanning start and target points, with an error ≤0.2mm; The attitude constraint is achieved by solving inverse kinematics and checking the effectiveness of joint space, so that the endpoint tangent vector is aligned with the point cloud normal vector, ensuring that the robot arm end posture fits the target surface. In addition, non-adjacent nodes are randomly selected, and if there is no collision when connecting straight lines, the intermediate nodes are deleted, thereby shortening the path length while satisfying the constraints.

8. The shallow ultrasound scanning path planning method based on the SVD-improved RRT algorithm as described in claim 7, characterized in that, In S105, the end-effector pose and camera observation pose of the ultrasound medical robot are obtained, and the calibration board is driven to perform hand-eye calibration, including: By solving the transformation relationship through matrix operations, the final hand-eye matrix calibration error is less than 0.8mm, achieving coordinate unification between the camera and the robot. The hand-eye transformation matrix B is solved based on the homogeneous transformation equation. ； If the average error between the planned path and the actual path in the x, y, and z directions based on the transformation matrix B is within 0.66 mm, 0.68 mm, and 0.82 mm respectively, then the shallow ultrasound scan path planning is completed by executing the scan trajectory.

9. A shallow ultrasound scanning path planning device based on SVD-improved RRT algorithm, characterized in that, The device includes a storage device and a processor, the storage device being used to store a computer program, and the processor running the computer program to cause the shallow ultrasound scanning path planning device to perform the shallow ultrasound scanning path planning method according to any one of claims 1-8.

10. The shallow ultrasound scanning path planning device based on the SVD-improved RRT algorithm as described in claim 9, characterized in that, The storage device includes: Point cloud processing module: performs point cloud data acquisition, 3D coordinate calculation, statistical filtering, and voxel downsampling operations; Path planning module: Performs SVD plane fitting, closed contour segmentation, and path point normal vector calculation; RRT optimization module: performs adaptive target point sampling, neighborhood reconnection and fifth-order B-spline smoothing algorithm; Control verification module: performs endpoint position / attitude constraint verification and path shortening strategy; The processor is configured to run the computer program to achieve: The Kinect camera interface is used to acquire shallow depth images in real time, which in turn drive the point cloud processing module to generate optimized point clouds. The path planning module is invoked to extract 16 equally divided path points and their normal vectors to construct the initial scan path; The RRT optimization module is activated to generate a smooth robotic arm trajectory, and the endpoint inverse kinematics solution parameters are loaded. The trajectory is output to the UR5e robotic arm through the robot control interface, and the control verification is triggered simultaneously to monitor the pose error and collision detection in real time.

Images