CN111127446A - A Plantar Pressure Image Partitioning Method for Gait Analysis - Google Patents

Abstract

The invention relates to a plantar pressure image partitioning method for gait analysis, comprising the following steps: acquiring plantar pressure data and performing preprocessing; clustering the preprocessed plantar pressure data to obtain the plantar pressure in the footprint area The image is integrated and projected to obtain the pressure center of the heel and the sole of the foot; the positioning model of the plantar pressure feature point is established, and the feature points of the plantar pressure image are located; The pressure images are partitioned, and the plantar pressure feature point location includes two stages: model training and feature point search; in the model training stage, a sufficient number of plantar pressure images are selected as the training set, and the exact positions of the feature points are manually marked to establish a foot print. The plantar pressure feature point location model; in the feature point search stage, the plantar pressure feature point location model is initialized according to the heel and sole pressure center, and then the exact location of the plantar pressure image feature point is found through the model search.

Description

A Plantar Pressure Image Partitioning Method for Gait Analysis

technical field

The invention relates to the field of gait analysis, in particular to a plantar pressure image partition method oriented to gait analysis.

Background technique

When the human body is standing still or walking dynamically, the sole of the foot will receive the reaction force from the ground, among which the reaction force perpendicular to the contact plane between the sole and the ground is the most obvious. This force is the sole pressure. When the human foot structure is diseased or dysfunctional, the plantar pressure distribution will also change dramatically. By dividing the plantar pressure into anatomical divisions, modern medical methods can compare the plantar pressure distribution of the healthy foot and the pathological foot with the pressure distribution of the same patient in different plantar pressure areas, which can assist doctors in diagnosing the cause of the patient's diseased foot. Designate rehabilitation treatment options.

The existing plantar pressure divisions are all semi-manual divisions that doctors or researchers rely on subjective experience to manually divide the plantar pressure data or apply a fixed template to the data acquisition software to preliminarily divide the plantar pressure data and then manually adjust them by the software user. These plantar pressure division methods It relies too much on human experience and subjective consciousness, and has the disadvantages of low efficiency and high cost, so it cannot be widely used, and it is difficult to be used for rapid partition processing of large quantities of plantar pressure data, and it cannot be used for gait analysis. Device's real-time plantar pressure zoning algorithm.

Therefore, it is necessary to develop an efficient, low-cost, and easy-to-code method for plantar pressure partitioning.

SUMMARY OF THE INVENTION

The technical problem of the present invention is to overcome the deficiencies of the prior art and provide a plantar pressure image partitioning method oriented to gait analysis, which can automatically select the footprint area from the plantar pressure image, and locate the feature points according to the method. Plantar pressure image partitioning has the advantages of high efficiency, low cost and rapid implementation.

The technical scheme adopted by the present invention is as follows: acquiring plantar pressure data and performing preliminary processing; clustering the pre-processed plantar pressure data to obtain a plantar pressure image in the footprint area; performing integral projection on the plantar pressure image to obtain a foot Heel and sole pressure center; establish a plantar pressure feature point location model to locate the feature points of the plantar pressure image; according to the relative positional relationship between each area of the foot and the plantar pressure feature point, the plantar pressure image is partitioned. Among them, the plantar pressure feature point location includes two stages: model training and feature point search. In the model training stage, a sufficient number of plantar pressure images are selected as the training set, and the exact positions of the feature points in the training samples are manually marked to establish the plantar. The pressure feature point location model is only executed once in the model training phase; in the feature point search phase, the plantar pressure feature point location model is initialized according to the heel and sole pressure center, and then the exact location of the plantar pressure image feature point is found through the model search. .

The specific implementation is as follows:

Step (1): Acquire plantar pressure data and perform preliminary processing, which specifically includes the following sub-steps:

Step (11): using an array pressure sensor to obtain plantar pressure data containing complete footprints in multiple adjacent frames (at least 3 frames) in the time domain;

Step (12): use the multi-frame plantar pressure data obtained in step (11) to perform time-domain mean filtering on the intermediate frame;

Step (13): performing maximum filtering processing and interpolation processing on the intermediate frame data obtained in step (12) in turn to obtain plantar pressure data with uniform size and equal spacing between rows and columns;

Step (2): Clustering the pre-processed plantar pressure data in step (1), locating the position of the footprint in the plantar pressure data, and obtaining a plantar pressure image of the footprint area, which specifically includes the following sub-steps:

Step (21): using the DBSCAN clustering algorithm to cluster the pre-processed plantar pressure data obtained in step (1) to obtain a plurality of pressure area blocks, and the footprint area may be composed of several pressure area blocks;

Step (22): Take the minimum circumscribed moment center point of each pressure area block obtained in step (21) as the center of the area block, and use the K-means clustering algorithm to aggregate all the pressure area blocks contained in each footprint into one. Large plantar pressure area block;

Step (23): extracting the pressure data of the footprint area obtained in step (22) to obtain a plantar pressure image including a complete footprint;

Step (3): perform integral projection on the plantar pressure image obtained in step (2) to obtain the heel and sole pressure center, which specifically includes the following sub-steps:

Step (31): take the long side direction of the minimum circumscribed moment of the plantar pressure image obtained in step (2) as the vertical axis, and the short side direction as the horizontal axis, and integrally project the plantar pressure image along the vertical axis to obtain a grayscale histogram;

Step (32): if the grayscale histogram obtained in step (31) contains a plurality of similar peak points, a sliding window with a length of N can be used to perform mean filtering on the grayscale histogram;

Step (33): because the sole is wider than the heel, the highest peak point of the filtered grayscale histogram is the center row of the sole, and the second highest peak point is the center row of the heel;

Step (34): Take 1/6 of the total number of rows of the plantar pressure image in the vertical middle row of the sole of the foot to obtain the sole area, and integrally project the image of the sole area along the horizontal axis to obtain the sole grayscale histogram, the sole grayscale histogram The highest peak point is the center row of the sole of the foot, the second highest peak point is the center row of the heel, the intersection point of the center row of the sole of the foot and the center row of the sole of the foot is the center point of the sole of the foot, and the intersection of the center row of the heel and the center row of the heel is the center point of the heel. ;

Step (4): establish a plantar pressure feature point location model, and locate the feature points of the plantar pressure image; it specifically includes two stages of model training and feature point search:

Step (41): in the model training stage, select a sufficient number of plantar pressure images as a training set, manually mark the exact positions of the feature points in the training samples, and establish a plantar pressure feature point positioning model, and this model training stage is only performed once;

Step (42): in the feature point search stage, initialize the plantar pressure feature point positioning model according to the heel and sole pressure center, and then find the exact position of the plantar pressure image feature point through the iterative model;

Further, the processing method of the model training stage is as follows: select N (N>200) plantar pressure images obtained in step (3) as a training set; manually mark each plantar pressure image accurately n (n> =6) The location of the feature points (the selection of the feature points includes the heel, the center point of the sole of the foot and the inflection point of the pressure contour of the sole of the foot, and the feature points should have obvious relative positional relationship with each area of the foot), the coordinates of the feature points after sorting A shape vector is formed in series in sequence, and the set of shape vectors obtained from N plantar pressure images is denoted as X _u , and the shape vector obtained from the i-th plantar pressure image is denoted as X _ui , X _ui =(x _i0 ,...,x _i(n-1) ,y _i0 ,...,y _i(n-1) ) ^T ,i=0,...,N-1, where x _ik , y _ik represent the i-th plantar pressure training image, respectively The horizontal and vertical coordinates of the feature point k, 0≤k<n, n is the number of feature points; the shape vector set X _{u is} used as the input for the establishment of the plantar pressure feature point positioning model, and the plantar pressure is established according to the shape vector set X _u The feature point positioning model is used when searching for plantar pressure feature points; the part of manually marking the exact position of the feature points in the training sample and establishing the plantar pressure feature point positioning model is only executed once when establishing the active shape model;

Further, the plantar pressure feature point location model is a template-based feature point location method, which can be an ASM model or an AAM model; first, the model can be normalized by the Procrustes method, that is, by translation, rotation, The scaling transformation operation aligns the shape vector set X _u to the same shape vector X on the basis of not changing the point distribution model, and the shape vector X is the feature point template; then, a local feature is constructed for each feature point; so far, The feature point positioning model is constructed;

Further, the aligning the shape vector set X _u to the same shape vector X includes the following steps:

对其他训练样本X_i进行平移、旋转和缩放，使所有样本尽可能与基准形状接近；训练样本X_i与形状基准

之间的接近程度使用欧式距离定义

得到变换后的形状向量为X′_i＝M(s,θ)[X_i]-t，其中，s为缩放尺度、θ为旋转角度、t平移向量；Step (411): take a certain training sample X _j as the shape reference

Translate, rotate and scale the other training samples X _i to make all samples as close to the reference shape as possible; the training samples X _i and the shape reference

The closeness between is defined using Euclidean distance

The transformed shape vector is obtained as X′ _i =M(s,θ)[X _i ]-t, where s is the scaling scale, θ is the rotation angle, and t is the translation vector;

并计算当前形状基准

与上次形状基准

之间的平移、旋转和缩放偏差；Step (412): Calculate the average shape of all transformed training samples X' _u as a new shape reference

and calculate the current shape datum

Same as last shape datum

translation, rotation, and zoom bias between;

Step (413): Iterative steps (411) and (412), if the deviation is less than the specified threshold or the iteration exceeds the specified maximum number of iterations, the iteration is stopped; the shape reference obtained at the last time is used as the average shape vector X after alignment of all training samples.

和方差S_i；首先，在第j(j＝0,1,…,N-1)个训练样本的第i个特征点两侧，沿垂直于该点前后两个特征点连线的方向上分别选择n个压力点，构成一个长度为2n+1的向量，对该向量所包含的压力值求导得到一个局部纹理g_ij，对训练集中所有样本同样操作可得到第i个特征点的N个局部纹理，计算均值

和方差

Further, the specific operation of constructing a local feature for each feature point is: calculating the average local texture of the feature point i (i=0,1,...,n-1) in the training sample

and variance S _i ; first, on both sides of the ith feature point of the jth (j=0,1,...,N-1) training sample, along the direction perpendicular to the line connecting the two feature points before and after the point Select n pressure points respectively to form a vector with a length of 2n+1, derive a local texture g _ij from the pressure value contained in the vector, and perform the same operation on all samples in the training set to obtain N of the i-th feature point local textures, compute the mean

and variance

Further, the feature point search stage includes initializing the model according to the heel and sole pressure center, and then finding the exact position of the feature point of the plantar pressure image through the iterative model; specifically, it includes the following sub-steps:

中足跟和脚掌压力中心(x′₁,y'₁)和(x'₂,y'₂)之间的平移、旋转、缩放偏差，对形状基准

进行变换得到模型的初始位置X_c；Step (421): Using the heel and sole pressure center as the benchmark, calculate the heel and sole pressure center (x ₁ , y ₁ ) and (x ₂ , y ₂ ) and the shape benchmark in the target plantar pressure image

Translation, rotation, scaling deviations between mid-heel and ball pressure centers (x' ₁ , y' ₁ ) and (x' ₂ , y' ₂ ), versus shape datum

Perform transformation to obtain the initial position X _c of the model;

之间的马氏距离，使得马氏距离最小的那个窗口的中心点作为特征点i的新位置。Step (422): search for the new position of each feature point in the plantar pressure image; first, overlay the initial position X _c of the active shape model on the plantar pressure image, and for the i-th feature point in the model, in the vertical direction It is centered on the line connecting the front and back two feature points, and m (m>n) pressure points are selected on each side, and the feature point i is added to form a search neighborhood with a length of 2m+1; let the length be n The window of the pressure point is slid within the search neighborhood, the local texture _gi for each window is calculated, and the local and average textures are calculated

The Mahalanobis distance between them, so that the center point of the window with the smallest Mahalanobis distance is used as the new position of the feature point i.

Step (423): Calculate the translation, rotation and scaling parameters of the shape vector of the initial position of the model and the shape vector after the updated position;

其中，x_newi、y_newi分别为新的形状向量X_new的第i个特征点的横坐标和纵坐标，x_ci、y_ci分别为原形状向量X_c的第i个特征点的横坐标和纵坐标。若

或循环达到最大次数，则搜索完成，至此，得到足底压力图像中每个特征点的准确位置；Step (424): Repeat steps (422) and (423) to calculate the distance _Dx between the new shape vector _Xnew and the original shape vector _Xc ,

Among them, x _newi and y _newi are the abscissa and ordinate of the ith feature point of the new shape vector X _new , respectively, and x _ci and y _ci are the abscissa and the ith feature point of the original shape vector X _c , respectively. Y-axis. like

Or when the cycle reaches the maximum number of times, the search is completed. At this point, the exact position of each feature point in the plantar pressure image is obtained;

Step (5): According to the relative positional relationship between each area of the foot and the plantar pressure feature points, combined with the feature points of the plantar pressure image obtained in step (4), the plantar pressure image is partitioned.

The present invention has the following beneficial effects:

(1) The plantar pressure partitioning method of the present invention can remove the contact noise, adhesion noise, network signal noise and acquisition circuit noise of the original plantar pressure data, and automatically extract the data from the soles of the feet through two clustering and minimum external moment algorithms. The extraction box in the pressure image selects the footprint area;

(2) The present invention adopts the method of multiple integral projection to detect the center point of the sole and the center point of the heel, which ensures the reliable positioning of the center point of the heel and the center point of the sole, and at the same time improves the accuracy of the model initialization in the search stage of the plantar pressure feature point The accuracy of plantar pressure feature point location and the accuracy of plantar pressure image division are improved;

(3) The plantar pressure partition method according to the present invention selects points with distinctive features, such as the heel center point, the sole center point, and the inflection point of the plantar pressure contour, as the feature points of the plantar pressure image, which is beneficial to the training sample set Establish a positioning model of plantar pressure feature points, and at the same time, there is an obvious relative positional relationship between the selected feature points and various areas of the foot, which is also conducive to dividing the plantar pressure image according to the feature points;

(4) The plantar pressure partition method according to the present invention is a method of partitioning the plantar pressure image by accurately locating the feature points of the plantar pressure image, and then based on the relative positional relationship between the feature points and each area of the foot , you can get accurate results of plantar pressure division;

(5) The model training phase of the plantar pressure feature point location model used in the present invention only needs to be performed once, and the trained model can be directly used in the plantar pressure feature point search phase, which ensures that the plantar pressure partition method can be highly efficient implement.

Description of drawings

In order to illustrate the technical solutions of the method of the present invention more clearly, the following briefly introduces the accompanying drawings used in the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, which are very important in the art. For those of ordinary skill, other drawings can also be obtained from these drawings without any creative effort.

Fig. 1 is the schematic flow sheet of the method of the present invention;

Fig. 2 is the schematic diagram of the integral projection of the plantar pressure image;

Fig. 3 is the sub-flow block diagram of the plantar pressure feature point positioning part in the embodiment of the present invention;

Fig. 4 is a schematic diagram of selection of plantar pressure feature points in an embodiment of the present invention;

5 is an effect diagram of the initial position X _c of the active shape model overlaid on the plantar pressure image in the embodiment of the present invention;

FIG. 6 is a schematic diagram of the result of the segmentation of the plantar pressure image in the embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

It should be noted that like numerals and letters refer to like items in the following figures, so once an item is defined in one figure, it does not require further definition and explanation in subsequent figures.

Example 1

As shown in FIG. 1 , the core of the present invention is to provide a method for determining a plantar pressure image for gait analysis, including the following steps:

Step (1): Acquire plantar pressure data and perform preliminary processing, which specifically includes the following sub-steps:

Step (11): using an array pressure sensor to obtain plantar pressure data containing complete footprints in multiple adjacent frames (at least 3 frames) in the time domain;

Step (12): use the multi-frame plantar pressure data obtained in step (11) to perform time-domain mean filtering on the intermediate frame;

Step (13): performing maximum filtering processing and interpolation processing on the intermediate frame data obtained in step (12) in turn to obtain plantar pressure data with uniform size and equal spacing between rows and columns;

Step (2): Clustering the pre-processed plantar pressure data in step (1), locating the position of the footprint in the plantar pressure data, and obtaining a plantar pressure image of the footprint area, which specifically includes the following sub-steps:

Step (21): using the DBSCAN clustering algorithm to cluster the pre-processed plantar pressure data obtained in step (1) to obtain a plurality of pressure area blocks, and the footprint area may be composed of several pressure area blocks;

Step (22): take the minimum circumscribed moment center point of each pressure area block obtained in step (21) as the center of the area block, and use the K-means clustering algorithm to gather the pressure area blocks into a footprint area;

Step (23): extracting the pressure data of the footprint area obtained in step (22) to obtain a plantar pressure image including a complete footprint;

Step (3): According to the method shown in FIG. 2, perform integral projection on the plantar pressure image obtained in step (2) to obtain the pressure center of the heel and the sole, which specifically includes the following sub-steps:

Step (31): take the long side direction of the minimum circumscribed moment of the plantar pressure image obtained in step (2) as the vertical axis, and the short side direction as the horizontal axis, and integrally project the plantar pressure image along the vertical axis to obtain a grayscale histogram;

Step (32): if the grayscale histogram obtained in step (31) contains a plurality of similar peak points, a sliding window with a length of N can be used to perform mean filtering on the grayscale histogram;

Step (33): because the sole is wider than the heel, the highest peak point of the filtered grayscale histogram is the center row of the sole, and the second highest peak point is the center row of the heel;

Step (34): Take 1/6 of the total number of rows of the plantar pressure image in the vertical middle row of the sole of the foot to obtain the sole area, and integrally project the image of the sole area along the horizontal axis to obtain the sole grayscale histogram, the sole grayscale histogram The highest peak point is the center row of the sole of the foot, the second highest peak point is the center row of the heel, the intersection point of the center row of the sole of the foot and the center row of the sole of the foot is the center point of the sole of the foot, and the intersection of the center row of the heel and the center row of the heel is the center point of the heel. ;

Step (4): as shown in Figure 3, a plantar pressure feature point location model is established, and the feature points of the plantar pressure image are located; it specifically includes two stages of model training and feature point search;

Step (41): In the model training stage, select N (N>200) plantar pressure images as the training set, manually mark the exact positions of the feature points in the training samples, and establish a plantar pressure feature point positioning model. The model is trained. stage is executed only once;

Step (42): in the feature point search stage, initialize the plantar pressure feature point positioning model according to the heel and sole pressure center, and then find the exact position of the plantar pressure image feature point through the iterative model;

Further, the processing method in the model training stage is: selecting N plantar pressure images obtained in step (3) as a training set; manually marking accurate n (n>=6) feature points of each plantar pressure image (The selection of feature points includes the heel, the center point of the sole of the foot and the inflection point of the pressure contour of the sole, and the feature point should have an obvious relative positional relationship with each area of the foot), as shown in Figure 4, as an example, this implementation In the example, 17 feature points are selected, namely: F1, F2, ..., F16, F17, including the heel center point F2 and the sole center point F1; the sorted coordinates of the feature points are connected in series to form a shape vector. The shape vector set obtained from the bottom pressure image is denoted as X _u , and the shape vector obtained from the i-th plantar pressure image is denoted as X _ui , X _ui =(x _i0 ,...,x _i(n-1) ,y _i0 , ...,y _i(n-1) ) ^T ,i=0,...,N-1, where x _ik and y _ik respectively represent the k-th feature point F _k (x _ik ) in the i-th plantar pressure training image , y _ik ) of the horizontal and vertical coordinates, 0≤k<n; the shape vector set X _{u is} used as the input for the establishment of the plantar pressure feature point positioning model, and the plantar pressure feature point positioning model is established according to the shape vector set X _u , for foot pressure It is used when searching for plantar pressure feature points; the parts of manually marking the exact positions of the feature points in the training samples and establishing the plantar pressure feature point positioning model are only executed once when establishing the active shape model;

Further, the plantar pressure feature point positioning model is a template-based feature point positioning method. As an example, an active shape model is used in this embodiment; first, the model can be normalized by Procrustes. The translation, rotation, and scaling transformation operations align the shape vector set X _u to the same shape vector X without changing the point distribution model, and the shape vector X is the feature point template; feature; at this point, the feature point positioning model has been constructed;

Further, the specific steps for aligning the shape vector set X _u to the same shape vector X are as follows:

选取X_u中第一个形状向量作为初始平均形状向量

step

Select the first shape vector in X _u as the initial average shape vector

将每个形状向量X_ui向平均形状向量

对齐，对齐过程中的变换向量记为T＝(scosθ,ssinθ,t_x,t_y)^T，其中，s为缩放尺度、θ为旋转角度、t_x为x轴平移向量、t_y为y轴平移向量，对齐后的形状向量记为X′_ui；step

Convert each shape vector X _ui to the mean shape vector

Alignment, the transformation vector in the alignment process is denoted as T=(scosθ,ssinθ,t _x ,t _y ) ^T , where s is the scaling scale, θ is the rotation angle, t _x is the x-axis translation vector, and ty is the _y -axis The translation vector, the aligned shape vector is denoted as X′ _ui ;

The operation of alignment is X′ _ui =X _ui T, wherein the calculation method of the transformation vector T is:

W为权重矩阵，计算方法为：in,

W is the weight matrix, and the calculation method is:

First, calculate the distance between feature points k(x _ik , y _ik ) and l(x _ik , y _ik ) in the ith shape:

其中方差V_Dkl为所有N个足底压力图像中特征点k(x_ik,y_ik)、l(x_ik,y_ik)距离D_ikl(i＝0,…,N-1)的方差；加权，值w_k表示特征点k的稳定程度；Then, calculate the weighted value of feature point k

where the variance V _Dkl is the variance of the distance D _ikl (i=0,...,N-1) between the feature points k (x _ik , y _ik ) and l (x _ik , y _ik ) in all N plantar pressure images; weighted , the value w _k represents the stability of the feature point k;

Finally, take w _k as the diagonal to make the diagonal matrix W, and the diagonal matrix is the weight matrix;

更新对齐后的所有形状向量X′_ui的平均形状向量，记为

step

Update the average shape vector of all shape vectors X′ _ui after alignment, denoted as

重复步骤

直到收敛或最大迭代次数，然后输出对齐后的形状向量，记为X；step

Repeat steps

Until convergence or the maximum number of iterations, then output the aligned shape vector, denoted as X;

The convergence judgment condition is: the transformation vector T between the two average shape vectors before and after the calculation, if the conditions |s-1|<0.001, |θ|<0.001π/180, |t|<0.01 are satisfied at the same time, then convergence;

和方差S_i；首先，在第j(j＝0,1,…,N-1)个训练样本的第i个特征点两侧，沿垂直于该点前后两个特征点连线的方向上分别选择h个压力点，构成一个长度为2h+1的向量，对该向量所包含的压力值求导得到一个局部纹理g_ij，对训练集中所有样本同样操作可得到第i个特征点的N个局部纹理，计算均值

和方差

Further, the specific operation of constructing a local feature for each feature point is: calculating the average local texture of the feature point i (i=0,1,...,n-1) in the training sample

and variance S _i ; first, on both sides of the ith feature point of the jth (j=0,1,...,N-1) training sample, along the direction perpendicular to the line connecting the two feature points before and after the point Select h pressure points respectively to form a vector with a length of 2h+1, derive a local texture g _ij from the pressure value contained in the vector, and perform the same operation on all samples in the training set to obtain N of the i-th feature point local textures, compute the mean

and variance

Further, the establishment of the active shape model includes PCA analysis, and the steps are as follows:

Step (PCA-1): Calculate the average shape vector of the aligned N shape vectors

Step (PCA-2): Calculate the covariance matrix of N shape vectors

Step (PCA-3): Calculate the eigenvalues λ _i of the covariance matrix and sort them from large to small, and the corresponding eigenvectors are denoted as p _i , i=0,1,...,2n-1;

Step (PCA-4): Select the top k largest eigenvalues, and form the corresponding eigenvectors into a principal component analysis matrix P=(p ₀ , p ₁ ,...,p _k-1 );

其中，b是一个k维形状参数，用来控制特征点的形状变化；此处将b约束为

Step (PCA-5): Build the active shape model as

Among them, b is a k-dimensional shape parameter, which is used to control the shape change of feature points; here, b is constrained to be

Further, the feature point search stage includes the following steps: initializing the active shape model according to the heel and sole pressure center, and then finding the exact position of the feature point of the plantar pressure image through the iterative model, and the specific steps are as follows:

Step (421): Initialize the active shape model through affine transformation; the coordinates of the heel and sole pressure center in the target plantar pressure image are respectively recorded as (x ₁ , y ₁ ) and (x ₂ , y ₂ ), the active shape model The coordinates of the center of heel and sole pressure are recorded as (x' ₁ , y' ₁ ) and (x' ₂ , y' ₂ ), respectively. First, calculate the scaling scale s and the rotation angle θ, and let the translation vectors t _x and t _y is 0, scale and rotate the active shape model to obtain a temporary shape, record the coordinates of the heel and sole pressure center of the temporary shape as (x″ ₁ , y″ ₁ ), and calculate the translation vector t _x , ty _y , then, let s=0, θ=0, translate the temporary shape to obtain the initial position X _c of the model;

Step (422): search for a new position of each feature point in the plantar pressure image; first, overlay the initial position X _c of the active shape model on the plantar pressure image, as shown in Figure 5, where point F _c 1 , F _c 2, ..., F _c 16, F _c 17, are all feature points in the initial position X _c . For the i-th feature point in the model, take it as the center in the direction perpendicular to the line connecting the two feature points before and after it, select m (m>h) pressure points on both sides, and add the feature point i to form a 2m-long pressure point +1 search neighborhood; let the window of length h pressure points slide within the search neighborhood, calculate the local texture g _i of each window, and calculate the Mahalanobis distance between the local texture and the average texture, so that the The center point of the window with the smallest distance is used as the new position of the feature point i;

Step (423): update attitude parameter; calculate the transformation vector T and deformation parameter b of the shape vector X _c of the initial position of the active shape model and the shape vector X _new after the updated position;

其中，x_newi、y_newi分别为新的形状向量X_new的第i个特征点的横坐标和纵坐标，x_ci、y_ci分别为原形状向量X_c的第i个特征点的横坐标和纵坐标。若

或循环达到最大次数，则搜索完成，至此，得到足底压力图像中每个特征点的准确位置；Step (424): Repeat steps (422) and (423) to calculate the distance _Dx between the new shape vector _Xnew and the original shape vector _Xc ,

Among them, x _newi and y _newi are the abscissa and ordinate of the ith feature point of the new shape vector X _new , respectively, and x _ci and y _ci are the abscissa and the ith feature point of the original shape vector X _c , respectively. Y-axis. like

Or when the cycle reaches the maximum number of times, the search is completed. At this point, the exact position of each feature point in the plantar pressure image is obtained;

Step (5): As shown in Figure 6, the points F _o 1, F _o 2, ..., F _o 16, and F _o 17 are all feature points F1, F2, ..., F16 in the plantar pressure image, The exact location of the F17. According to the relative positional relationship between each area of the foot and the plantar pressure feature points, combined with the feature points of the plantar pressure image obtained in step (4), the plantar pressure image is divided; The bottom pressure image is divided into six areas: toe area, sole area, medial medial foot, medial lateral foot, medial heel and lateral heel; the specific steps are: the envelope surface of F _o 3-F _o 17 feature points is: The entire footprint area; the envelope surface of F _o 3-F _o 4-F _o 5-F _o 16-F _o 17 feature points is the heel area, and the connecting line between F _o 3 and F _o 2 feature points can be Divide the heel area into the medial area of the heel and the lateral area of the heel; the envelope surface of the feature points F _o 5-F _o 6-F _o 15-F _o 16 is the midfoot area, F _o 1 and F _o The line connecting the feature points No. 2 can further divide the midfoot area into the midfoot area and the midfoot area; F _o 7-F _o 8-F _o 9-F _o 13-F _o 14-F _o 15 Features The envelope surface of the point is the sole area; the envelope surface of the feature points F _o 9-F _o 10-F _o 11-F _o 12-F _o 13 is the toe area.

Claims (10)

1. a plantar pressure image partitioning method for gait analysis, is characterized in that, comprises the steps: Step (1): Acquire plantar pressure data and perform preliminary processing; Step (2): clustering the pre-processed plantar pressure data in step (1), locating the position of the footprint in the plantar pressure data, and obtaining a plantar pressure image in the footprint area; Step (3): perform integral projection on the plantar pressure image obtained in step (2) to obtain the heel and sole pressure center; Step (4): establish a plantar pressure feature point positioning model, and locate the feature points of the plantar pressure image; Step (5): According to the relative positional relationship between each area of the foot and the plantar pressure feature points, combined with the feature points of the plantar pressure image obtained in step (4), the plantar pressure image is partitioned. 2. The plantar pressure image partitioning method for gait analysis according to claim 1, wherein the specific implementation process of the step (1) comprises: Step (11): using an array pressure sensor to obtain plantar pressure data containing complete footprints in multiple adjacent frames (at least 3 frames) in the time domain; Step (12): use the multi-frame plantar pressure data obtained in step (11) to perform time-domain mean filtering on the plantar pressure data of the intermediate frame; Step (13): Perform maximum filtering processing and interpolation processing on the plantar pressure data of the intermediate frame obtained in step (12) in turn to obtain pre-processed plantar pressure data with uniform size and equal spacing between rows and columns. 3. The plantar pressure image partitioning method for gait analysis according to claim 1, wherein the specific process of the step (2) comprises: Step (21): use the DBSCAN clustering algorithm to cluster the pre-processed plantar pressure data obtained in step (1) to obtain a plurality of pressure area blocks, and the footprint area is composed of several pressure area blocks; Step (22): Take the center point of the minimum circumscribed moment of each pressure area block in the multiple pressure area blocks obtained in step (21) as the center of the pressure area block, and use the K-means clustering algorithm to cluster the pressure area blocks footprint area; Step (23): extracting the pressure data of the footprint area obtained in step (22) to obtain a plantar pressure image including a complete footprint. 4. The plantar pressure image partitioning method for gait analysis according to claim 1, wherein the specific process of the step (3) comprises: Step (31): take the long side direction of the minimum circumscribed moment of the plantar pressure image obtained in step (2) as the vertical axis, and the short side direction as the horizontal axis, and integrally project the plantar pressure image along the vertical axis to obtain a grayscale histogram ; Step (32): if the grayscale histogram obtained in step (31) contains a plurality of similar peak points, a sliding window with a length of N is used to perform mean filtering on the grayscale histogram; Step (33): based on the width of the sole of the foot than the heel, the highest peak point of the filtered grayscale histogram is the center row of the sole of the foot, and the second highest peak point is the center row of the heel; Step (34): taking 1/6 of the total number of rows of the plantar pressure images on the vertical center row of the sole of the foot on each of the upper and lower sides to obtain the sole area, and integrally projecting the image of the sole area along the horizontal axis to obtain the sole gray histogram, the sole gray histogram The highest peak point in the figure is the center column of the sole of the foot, the second highest peak point is the center column of the heel, the intersection of the center row of the sole of the foot and the center column of the sole of the foot is the center of the sole of the foot, and the intersection of the center row of the heel and the center column of the heel is the center of the heel. point. 5. the plantar pressure image partitioning method for gait analysis according to claim 1, is characterized in that: described step (4) realization process: comprise two stages of model training and feature point search: Step (41): In the model training stage, select N plantar pressure images as the training set, N>200; manually mark the exact positions of the feature points in the training samples, and establish a plantar pressure feature point positioning model, the model training stage only execute once; Step (42): In the feature point search stage, initialize the plantar pressure feature point positioning model according to the heel and sole pressure center, and then find the exact position of the plantar pressure image feature point through the iterative model. 6. The plantar pressure image partition method for gait analysis according to claim 5, characterized in that: the processing method in the model training stage is: selecting N plantar pressure images obtained in step (3) as training Set; manually mark the exact positions of n feature points for each plantar pressure image, n is the number of feature points, n>=6, and the selection of feature points includes the heel, the center point of the sole and the contour of the plantar pressure. Inflection point, feature point should have obvious relative positional relationship with each area of the foot, the sorted coordinates of the feature points are connected in series to form a shape vector, the shape vector set obtained from N plantar pressure images is denoted as X _u , where the i-th The shape vector obtained from the plantar pressure image is denoted as X _ui , X _ui =(x _i0 ,...,x _i(n-1) ,y _i0 ,...,y _i(n-1) ) ^T ,i=0, ...,N-1, where x _ik , y _ik represent the horizontal and vertical coordinates of the k-th feature point F _k (x _ik , y _ik ) in the ith plantar pressure image, 0≤k<n; the shape vector Set X _u as the input of establishing the plantar pressure feature point positioning model part, build the plantar pressure feature point positioning model according to the shape vector set X _u , use when searching for the plantar pressure feature point, and mark the feature points in the training sample by hand. The exact location of the plantar pressure feature point location model and the establishment of the plantar pressure feature point location model parts are only performed once when the active shape model is established. 7. The plantar pressure image partitioning method for gait analysis according to claim 6, wherein the plantar pressure feature point location model is a template-based feature point location, using ASM model or AAM model Establishment, the establishment method is: first, through the Procrustes normalization method, that is, through translation, rotation, scaling transformation operations, without changing the point distribution model, align the shape vector set X _u to the same shape vector X, the shape The vector X is the feature point template; then, a local feature is constructed for each feature point, and the feature point positioning model is completed. 8. The plantar pressure image partitioning method for gait analysis according to claim 7, characterized in that: said aligning the shape vector set X _u to the same shape vector X comprises the following steps: Step (411): take a certain training sample X _j as the shape reference

Translate, rotate and scale the other training samples X _i so that all samples are close to the reference shape; the training samples X _i are similar to the shape reference

The closeness between is defined using Euclidean distance

The transformed shape vector is obtained as X′ _i =M(s,θ)[X _i ]-t, where s is the scaling scale, θ is the rotation angle, and t is the translation vector; Step (412): Calculate the average shape of all transformed training samples X' _u as a new shape reference

and calculate the current shape datum

Same as last shape datum

translation, rotation, and zoom bias between; Step (413): Iterative steps (411) and (412), if the deviation is less than the specified threshold or the iteration exceeds the specified maximum number of iterations, the iteration is stopped; the shape reference obtained last time is used as the average shape vector X after alignment of all training samples. 9. The plantar pressure image partitioning method for gait analysis according to claim 7, wherein the specific operation of constructing a local feature for each feature point is: calculating the average local value of the feature point i in the training sample texture

and variance S _i , i=0,1,...,n-1; firstly, on both sides of the i-th feature point of the j-th training sample, j=0,1,...,N-1, along the perpendicular to the Select h pressure points in the direction of the line connecting the two feature points before and after the point to form a vector with a length of 2h+1, and derive a local texture g _ij from the pressure value contained in the vector. For all samples in the training set The same operation is used to obtain N local textures of the i-th feature point, and the mean value is calculated.

and variance

10. The gait analysis-oriented plantar pressure image partitioning method according to claim 5, wherein the feature point search stage comprises: initializing the model according to the pressure center of the heel and the sole, and then finding the foot pressure through the iterative model. The exact position of the feature point of the bottom pressure image; the specific steps are: Step (421): Using the heel and sole pressure center as the benchmark, calculate the heel and sole pressure center (x ₁ , y ₁ ) and (x ₂ , y ₂ ) and the shape benchmark in the target plantar pressure image

Translation, rotation, scaling deviation between mid-heel and ball pressure centers (x' ₁ , y' ₁ ) and (x' ₂ , y' ₂ ), versus shape datum

Perform transformation to obtain the model initial position X _c ; Step (422): search for the new position of each feature point in the plantar pressure image; first, overlay the initial position X _c of the active shape model on the plantar pressure image, and for the i-th feature point in the model, in the vertical direction It is centered on the line connecting the front and rear feature points, and m pressure points are selected on each side, m>h, and the feature point i is added to form a search neighborhood with a length of 2m+1; let the length be h pressure points A window of points slides within the search neighborhood, computes the local texture g _i for each window, and computes the local and average textures

The Mahalanobis distance between, so that the center point of the window with the smallest Mahalanobis distance is used as the new position of the feature point i; Step (423): calculate the translation, rotation and scaling parameters of the shape vector of the model initial position X _c and the shape vector after the updated position; Step (424): Repeat steps (422) and (423) to calculate the distance _Dx between the new shape vector _Xnew and the original shape vector _Xc ,

Among them, x _newi and y _newi are the abscissa and ordinate of the ith feature point of the new shape vector X _new , respectively, and x _ci and y _ci are the abscissa and the ith feature point of the original shape vector X _c , respectively. ordinate, if

Or when the cycle reaches the maximum number of times, the search is completed, and the search is completed, and the exact position of each feature point in the plantar pressure image is obtained.

Images