Search

CN-121977606-A - Pose estimation method, device and computer program product of magnetic control capsule

CN121977606ACN 121977606 ACN121977606 ACN 121977606ACN-121977606-A

Abstract

The invention discloses a pose estimation method, a pose estimation device and a computer program product of a magnetic control capsule, belonging to the technical field of magnetic micro-robot positioning, aiming at solving the problems of weak coupling processing capacity of multiple magnetic sources or lack of physical constraint in the pose estimation method. The method comprises the steps of processing three physical parameters, namely three-axis magnetic field components, coordinates of the magnetic sensors and intervals among the magnetic sensors, extracting features of the most basic physical parameters from data of one-dimensional independent magnetic sensors, data change among two-dimensional magnetic sensors and three-dimensional magnetic field vector angles by the obtained physical feature data package, analyzing and classifying the features according to the angles of geometric space, and outputting pose prediction data with physical consistency only by using a pose residual neural network for prediction, wherein the pose prediction data with physical consistency can be output based on the pose residual neural network and the magnetic sensor arrays, the multi-magnetic source coupling processing capability is strong, and the accurate positioning requirement under a complex magnetic environment is met.

Inventors

  • LU SHICHENG
  • Ge Wanning
  • FAN XINJIAN
  • Zhang Xibu

Assignees

  • 苏州大学

Dates

Publication Date
20260505
Application Date
20260407

Claims (10)

  1. 1. A pose estimation method of a magnetic control capsule is characterized by comprising the following steps of, Acquiring triaxial magnetic field components of each magnetic sensor (2) induction magnetic control capsule (4) in a magnetic sensor array, coordinates of each magnetic sensor (2) and intervals among different magnetic sensors (2); obtaining a physical characteristic data packet according to the triaxial magnetic field component of each magnetic sensor (2), the coordinates of each magnetic sensor (2) and the intervals among different magnetic sensors (2), wherein the physical characteristic data packet comprises original magnetic field data, physical field intensity characteristics, spatial gradient characteristics, magnetic field direction characteristics, statistical distribution characteristics and relative intensity characteristics; inputting the physical characteristic data packet into a trained pose residual neutral network to obtain the estimated pose of the magnetic control capsule (4).
  2. 2. The method for estimating the pose of a magnetically controlled capsule according to claim 1, wherein the physical characteristic data packet comprises 100-dimensional data, and the magnetic sensor array is a 3 x 3 distributed matrix; The raw magnetic field data is 27-dimensional in total and comprises magnetic sensors (2) An axial magnetic field component; the physical field intensity characteristics are 9-dimensional in total, and the magnetic field intensity module length of each magnetic sensor (2) is included, and is obtained through the following calculation: ; the spatial gradient feature is 20-dimensional in total, and comprises the magnetic field intensity change rate between every two magnetic sensors (2), wherein the magnetic field intensity change rate is obtained through the following formula: , There are 20 different combinations; The magnetic field direction characteristics are 27-dimensional in total and comprise each magnetic sensor (2) A unit direction vector of the axis, the magnetic sensor (2) The unit direction vector of the shaft is obtained by the following calculation: , ; Wherein, the 、 And Respectively the first The magnetic sensor is at The component of the magnetic field on the shaft, 、 、 The counting serial numbers of the magnetic sensors are respectively, 、 、 , 、 And Respectively the first 、 、 The magnetic field intensity of each magnetic sensor is in a mode length, and , In order to perform the operation of taking the mould, And From origin to the first of the magnetic sensor array Magnetic sensor and the first The distance of the individual magnetic sensors, Is the first Magnetic sensor and the first The rate of change of the magnetic field strength between the individual magnetic sensors, Is the first Of individual magnetic sensors The unit directional vector of the axis, Is a coefficient for controlling the numerical stability.
  3. 3. The method for estimating the pose of the magnetic control capsule according to claim 2, characterized in that the magnetic sensor (2) is numbered in the same direction line by line starting from the edge of the magnetic sensor mounting plane (1); Respectively of 20 combinations of (1,2)、(1,3)、(1,4)、(1,5)、(1,6)、(1,7)、(1,8)、(1,9)、(2,3)、(2,4)、(2,5)、(2,6)、(2,7)、(2,8)、(2,9)、(3,4)、(3,5)、(3,6)、(3,7)、(3,8).
  4. 4. The method for estimating the pose of a magnetically controlled capsule according to claim 2, wherein, The statistical distribution characteristics are 8-dimensional in total and comprise a mean value, a standard deviation, a skewness, a peak magnetic field intensity, the strongest sensor pose and an estimated distance, and the statistical distribution characteristics are obtained through the following calculation: ; the relative intensity features are 9-dimensional in total, and are obtained through the following calculation: ; Wherein, the Is the mean value of the two values, Is the standard deviation of the two-dimensional image, In order to be the degree of deviation, For the peak magnetic field strength, In order to operate at the maximum value, 、 And Respectively is The strongest sensor pose on the axis, Is an estimated constant for the magnetic dipole model, In order to estimate the distance it is possible, As a characteristic of the relative intensity of the light, Is the first The relative intensities of the individual magnetic sensors are, 、 And The magnetic field intensity mode lengths of the 1 st, 2 nd and 9 th magnetic sensors respectively, 、 And The relative intensities of the 1 st, 2 nd and 9 th magnetic sensors are shown, Transpose the symbols for the matrix.
  5. 5. The method for estimating the pose of the magnetically controlled capsule according to claim 2, wherein the pose residual neural network trains the pose training set and the radian value in the pose verification set from the following points Is mapped to the value interval of (a) Is a value interval of the number; the method for obtaining the estimated pose of the magnetic control capsule (4) further comprises the steps of carrying out recovery mapping on an estimated pose part in the estimated pose of the magnetic control capsule (4) to obtain a recovered estimated pose, wherein the recovered estimated pose is obtained through the following formula: , Wherein, the In order to recover the estimated pose, Is an estimated pose part in the estimated poses of the magnetic control capsule.
  6. 6. The method for estimating the pose of a magnetically controlled capsule according to claim 4, wherein said inputting said physical feature data packet into a trained pose residual neural network to obtain an estimated pose of said magnetically controlled capsule (4) comprises, The trained pose residual neural network comprises a feature projection layer, a grid remolding layer, a rotary and other denaturation convolution layer, a global feature extraction layer, a pose prediction head and a distance constraint branch; The feature projection layer maps the physical feature data packet from 100 dimensions to 256 dimensions according to a set loss rate to obtain a high-dimensional feature vector; The grid remodeling layer distributes each dimension of the high-dimensional feature vector to each magnetic sensor (2) according to an array distribution structure of the magnetic sensor array, and each magnetic sensor (2) is distributed with 32 dimensions to obtain a distributed feature vector; The rotation isodenaturing convolution layer carries out group convolution on the assigned feature vectors of all the magnetic sensors (2) to obtain 256-dimensional high-dimensional space vectors corresponding to each magnetic sensor (2); the global feature extraction layer extracts 256-dimensional high-dimensional space vectors of all the magnetic sensors (2) to obtain global features; The pose prediction head obtains the estimated pose of the magnetic control capsule (4) according to the global feature, and meanwhile the distance constraint branch outputs the prediction reliability of the pose prediction head according to the global feature, wherein the prediction reliability is used for adjusting the output weight of the pose prediction head.
  7. 7. The method for estimating the pose of a magnetically controlled capsule according to claim 6, wherein, The rotation equal-variability convolution layer comprises a first convolution layer, a second convolution layer and a third convolution layer, and the group convolution process of the rotation equal-variability convolution layer is as follows: the first convolution layer ascends the allocated feature vector from 32 dimensions to 64 dimensions to obtain a feature vector after first convolution, and the training random discarding rate of the first convolution layer is 0.4; The second convolution layer ascends the dimension of the feature vector after the first convolution from 64 dimensions to 128 dimensions to obtain the feature vector after the second convolution, and the training random discarding rate of the second convolution layer is 0.4; And the third convolution layer ascends the dimension of the feature vector after the second convolution from 128 dimensions to 256 dimensions to obtain 256-dimension high-dimensional space vectors, and the training random discarding rate of the third convolution layer is 0.3.
  8. 8. The method for estimating the pose of a magnetically controlled capsule according to any one of claims 1 to 7, wherein said inputting said physical feature data packet into a trained pose residual neural network comprises, Performing Z-score standardization on the physical characteristic data packet to obtain a standardized characteristic data packet, and inputting the standardized characteristic data packet into a trained pose residual neutral network, wherein the standardized characteristic data packet is obtained through the following calculation: , Wherein, the In order to standardize the feature data package, In order to be a data packet of a physical characteristic, Is the mean value of the data packet of the physical characteristic, Is the standard deviation of the physical characteristic data packet, The average value is obtained.
  9. 9. A pose estimation device of a magnetic control capsule, characterized in that the pose estimation device comprises: a magnetic sensor mounting plane (1) for mounting the magnetic sensors (2) in a 3 x 3 layout; -an analyzer (3), the analyzer (3) being in signal connection with a plurality of the magnetic sensors (2), respectively, the analyzer (3) being adapted to perform the pose estimation method of the magnetically controlled capsule according to any one of claims 1 to 8.
  10. 10. A computer program product comprising computer programs/instructions which, when executed by a processor, implement the steps of the method for estimating the pose of a magnetically controlled capsule according to any of claims 1 to 8.

Description

Pose estimation method, device and computer program product of magnetic control capsule Technical Field The invention relates to a pose estimation method, a pose estimation device and a computer program product of a magnetic control capsule, belonging to the technical field of magnetic micro-robot positioning. Background After the magnetic micro-robot enters a sensitive area, such as a human body, the magnetic micro-robot needs to be positioned to make reference for control instructions and play a role in ensuring safety. The existing pose estimation method of the magnetic control capsule is divided into two types, namely, the first method is based on a magnetic dipole physical model, the position and the direction of a magnetic source are calculated by inversion of a nonlinear optimization algorithm through measuring space magnetic field distribution, the method has the advantages of definite physical significance, the defect of weak coupling processing capacity of multiple magnetic sources and difficulty in coping with accurate positioning requirements under a complex magnetic environment, and the second method is based on a data-driven intelligent positioning method, wherein the method is used for establishing nonlinear mapping from magnetic field signals to poses only through a machine learning algorithm so as to further realize the prediction of the poses of the magnetic source, but the method encounters development bottleneck, physical consistency is lost, and a pure data driving method lacks physical constraint and possibly leads to position calculation results which do not accord with physical laws. Therefore, the main-stream pose estimation method of the magnetic control capsule has the problems of weak coupling processing capacity of multiple magnetic sources or lack of physical constraint. Disclosure of Invention The application aims to overcome the defects in the prior art and provide a pose estimation method, a pose estimation device and a computer program product of a magnetic control capsule with weak multi-magnetic source coupling processing capability and physical consistency. In order to achieve the above purpose, the application is realized by adopting the following technical scheme: in a first aspect, the present application provides a method for estimating the pose of a magnetically controlled capsule, comprising, Acquiring triaxial magnetic field components of each magnetic sensor induction magnetic control capsule in a magnetic sensor array, coordinates of each magnetic sensor and intervals among different magnetic sensors; Obtaining a physical characteristic data packet according to the triaxial magnetic field component of each magnetic sensor, the coordinates of each magnetic sensor and the intervals among different magnetic sensors, wherein the physical characteristic data packet comprises original magnetic field data, physical field intensity characteristics, spatial gradient characteristics, magnetic field direction characteristics, statistical distribution characteristics and relative intensity characteristics; and inputting the physical characteristic data packet into a trained pose residual neutral network to obtain the estimated pose of the magnetic control capsule. In the first aspect of the present application, the physical characteristic data packet includes 100-dimensional data, and the magnetic sensor array is a matrix distributed in 3×3; the raw magnetic field data is 27-dimensional in total and comprises magnetic sensors An axial magnetic field component; The physical field intensity characteristics are 9-dimensional in total and comprise the magnetic field intensity module length of each magnetic sensor, and the magnetic field intensity module length is obtained through the following calculation: ; The spatial gradient characteristics are 20-dimensional in total, and comprise the magnetic field intensity change rate between every two magnetic sensors, wherein the magnetic field intensity change rate is obtained through the following calculation: , There are 20 different combinations; the magnetic field direction characteristics are 27-dimensional in total and comprise each magnetic sensor Unit direction vector of axis, the magnetic sensorThe unit direction vector of the shaft is obtained by the following calculation: , ; Wherein, the 、AndRespectively the firstThe magnetic sensor is atThe component of the magnetic field on the shaft,、、The counting serial numbers of the magnetic sensors are respectively,、、,、AndRespectively the first、、The magnetic field intensity of each magnetic sensor is in a mode length, and,In order to perform the operation of taking the mould,AndFrom origin to the first of the magnetic sensor arrayMagnetic sensor and the firstThe distance of the individual magnetic sensors,Is the firstMagnetic sensor and the firstThe rate of change of the magnetic field strength between the individual magnetic sensors,Is the firstOf individual magnetic sensorsThe u