Height estimation method based on plantar pressure

Abstract

The present invention provides a height estimation method based on plantar pressure, which comprises the steps of collecting plantar pressure image samples and preprocessing them; establishing and training a neural network model; inputting the preprocessed image samples into the trained neural network model to calculate plantar pressure characteristic values; establishing a multiple linear regression model of plantar pressure characteristic values ​​and height and continuing to calculate plantar pressure characteristic values ​​by using the neural network model, estimating height by using the multiple linear regression model, etc. The method for obtaining plantar pressure characteristic values ​​of the present invention is scientific and reasonable, with high precision, and the accuracy of image interpretation by using a deep learning neural network is higher and higher as the number of interpretations increases, and the significance level and confidence of the multiple linear regression model are further improved.

Description

Technical Field

[0001] The invention belongs to the technical field of criminal investigation, and in particular relates to a height estimation method based on plantar pressure. Background Art

[0002] Using footprints at the crime scene to estimate the height of a suspect is one of the important tasks in the field of criminal investigation technology. Forensic anthropology has proven that a person's height is correlated with some special parts of the body, such as hands and feet, upper limbs and lower limbs. This provides a theoretical basis for using anatomical and mathematical ideas to estimate height through footprint analysis. Over the years, many experts have achieved fruitful research results in this field. The specific method is to analyze and explain the close relationship between the local characteristic measurement values ​​of footprints and height, establish a multivariate linear regression model that reaches the significance level, collect a large amount of footprint information of the human body in normal walking state, measure the 6 characteristic points in the plantar area and heel area, extract the local characteristic values ​​of the footprint, use the T test criterion to eliminate outliers in the collected data, and perform correlation analysis and model analysis on the sorted data to obtain regression parameters when the significance reaches the required level. Each set of parameters corresponds to a multivariate linear regression equation, and the equation is used as a footprint height estimation model. A large number of tests have shown that the model has an accuracy rate of more than 95% in measuring the height of people aged 20 to 35, about 80% in the age group of 35 to 45, and about 75% in people over 45 years old. Through the changing trend of the accuracy rate, it is not difficult to see that with the increase of age, a person's height will show a trend of nonlinear change with the footprint characteristic value. This trend change is in line with the natural law of human growth. It seems that the evolution and iteration of this calculation model has a ceiling that is not easy to break through. In order to break through this technical bottleneck, the present invention has tried a new technical route to find a breakthrough.

[0003] At present, all footprint estimation methods use the distance between the feature points on barefoot prints or sock prints as the input value of the independent variable of the multivariate linear regression model, and then estimate the height according to the model used. Two questions will inevitably arise here: First, the footprints on the scene are often fuzzy and overlapping incomplete footprints, which brings many uncertain factors to the determination of the precise location of the feature points, so how to remove the influence of these uncertain factors; second, the footprint information is mostly presented in the form of fuzzy irregular surfaces. Is there a tool that can obtain more accurate surface parameter information by interpreting the surface, and make a more accurate calculation of the height through these surface information? With these two questions, the present invention attempts a new interpretation method.

[0004] Based on the same theoretical basis as estimating height through the measurement of footprint feature points, it is not difficult to find that the various characteristic areas on a person's footprint can more accurately reflect the correlation with height. When the plantar image is gray-processed, these areas will be converted into concentric circle areas with different gray levels, and the size of the circumscribed frame of each area with the same gray level corresponds to the pressure value of the plantar of the human body. These areas are relatively independent of each other due to the characteristics of the plantar bone structure. At the same time, each area has a close relationship with the height. This provides a theoretical basis for using the pressure values ​​of these plantar characteristic areas as regression parameters of the multivariate linear regression model for regression analysis to estimate the height. When performing such an operation, the most critical and most difficult technical problem to solve is how to accurately judge the characteristic values ​​of the plantar pressure characteristic areas. Fortunately, in recent years, with the rapid development of artificial intelligence and the substantial improvement of computer hyperthreading computing power, especially the leapfrog development of GCU in the field of computing, the introduction of artificial intelligence neural network algorithms with deep learning capabilities into the intelligent interpretation of plantar pressure characteristic areas provides technical support for solving the technical problem of difficult extraction of plantar pressure values, which also represents the future development direction of this field. Summary of the invention

[0005] The purpose of the present invention is to introduce an artificial intelligence neural network algorithm with deep learning ability into the interpretation of the plantar pressure characteristic area, extract the pressure value parameters of the plantar characteristic area, and estimate the height through the analysis of the multivariate linear regression model of the plantar characteristic area pressure value and the height.

[0006] Specifically, the present invention provides a height estimation method based on plantar pressure, the height estimation method comprising the following steps:

[0007] Step 1. Collect plantar pressure image samples and preprocess them;

[0008] Step 2. Build and train the neural network model;

[0009] Step 3. Input the preprocessed image samples into the trained neural network model to calculate the plantar pressure characteristic value;

[0010] Step 4. Establish a multiple linear regression model between plantar pressure characteristic values ​​and height;

[0011] Step 5. Continue to use the neural network model to calculate the plantar pressure characteristic value and estimate the height through the multivariate linear regression model.

[0012] Furthermore, the collecting of plantar pressure image samples and preprocessing in step 1 includes the following steps:

[0013] Step 1.1. Define the plantar pressure feature areas in the plantar pressure image sample of an adult in walking state, including: the first metatarsal area, the second to third metatarsal areas, the fourth to fifth metatarsal areas, the heel area, the first phalanx area, the second to fifth phalanx areas and the arch area;

[0014] Step 1.2. Screen the collected plantar pressure image samples according to the expression ability of the plantar pressure characteristic area on the pressure characteristics, and eliminate the plantar pressure images that lack the first metatarsal area, the first phalanx area or the heel area in the walking state;

[0015] Step 1.3. Remove the additional information such as the scale, watermark and grid line in the screened plantar pressure image that is not suitable for neural network interpretation;

[0016] Step 1.4. For the plantar pressure image whose footprint direction and footprint position are not in the center of the image, rotation correction and footprint centering operations are performed;

[0017] Step 1.5. Using the point 10 pixels outward from the most prominent point on the upper and lower sides of the footprint as the tangent point, crop the image into a rectangular image of uniform size and enhance the resolution.

[0018] Furthermore, in step 2, a neural network model is established and trained, and the VGG network is used to have strong robustness in image interpretation and a neural network model is established based on the VGG16 network:

[0019] Step 2.1. The VGG16 network has a total of 16 layers, including 13 convolutional layers and 3 fully connected layers. The 13 convolutional layers are divided into 5 convolutional blocks. The first convolutional block contains 2 3*3*64 convolutional layers, the second convolutional block contains 2 3*3*128 convolutional layers, the third convolutional block contains 3 3*3*256 convolutional layers, the fourth convolutional block contains 3 3*3*256 convolutional layers, and the fifth convolutional block contains 3 3*3*256 convolutional layers. The first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, and the fifth convolutional block are all connected to a maximum pooling layer. The maximum pooling layer connected to the fifth convolutional block is connected to 3 fully connected layers. The last fully connected layer among the 3 fully connected layers is connected to the Softmax layer. The output of each fully connected layer is used as the input of the next fully connected layer.

[0020] Step 2.2. Input layer: Input a 224*224*3 local context area image to the VGG16 network. The input image is convolved with the convolution block to obtain multiple feature maps corresponding to each convolution block.

[0021] Step 2.3. The activation function performs nonlinear processing on multiple feature maps corresponding to each convolution block. The activation function uses the ReLU function. After the input image passes through the convolution and pooling layers, the feature map obtained by the last fully connected layer is further processed by the Softmax layer to obtain the classification result, which is output through the output layer.

[0022] Step 2.4. In view of the feature that the plantar pressure image contains few color difference pixels, based on the VGG16 network model, a batch normalization layer (BatchNormalization, BN) is introduced after each pooling layer to accelerate the training of the neural network model. The neural network model training input sample size is set to be no less than 4000, the batch size is 128 images per batch, the number of iterations is no less than 350 times, and the learning utilization rate is 0.0001.

[0023] Furthermore, the step 3 of calculating the plantar pressure characteristic value includes the following steps:

[0024] Step 3.1. Perform denoising on the preprocessed image samples. First, use a bilateral mean filter to reduce the noise interference in the spatial domain of the image while maintaining the edge texture blur information of the plantar image. Then, use a median filter to eliminate the salt-and-pepper noise in the image while retaining the sharpness of the edge of the plantar image.

[0025] Step 3.2. Convert the preprocessed image samples into grayscale images and mark the plantar pressure feature area. The marked area contains the fuzzy texture information of the edge of the plantar pressure feature area and is tangent to the arc with the largest curvature of the edge of the plantar pressure feature area.

[0026] Step 3.3. Perform histogram equalization on the marked area, scan all pixels in the marked area, and calculate the normalized histogram H of the marked area. The calculation process is to first list the differential histogram H(r k ) expression, and then the differential histogram H(r k ) integral, the calculation process is expressed as follows (1) and (2):

[0027]

[0028] H′(i)=∑ 0≤j≤i H(j)……(2),

[0029] In the above formulas (1) and (2), n is the total number of images, k is the kth grayscale level in the image, and n k is the number of gray levels, i represents the highest gray level in the region, and j represents the gray level variable in the region;

[0030] Finally, the image is transformed, which is expressed as the following formula (3):

[0031] dst(x,y)=H′(src(x,y))……(3),

[0032] In the above formula (3), dst(x, y) is the output image, and src(x, y) represents the input image;

[0033] Step 3.4. Use the neighborhood block Gaussian weighted method to perform local adaptive binary threshold segmentation on the image of the marked area, divide the image of the marked area into the focus area and the remaining area, and extract the focus area for further analysis;

[0034] Step 3.5. Perform median filtering on the area of ​​interest again. After obtaining an interference-free image, perform 3×3 kernel erosion first, and then perform 12×12 kernel dilation to complete the morphological operation of the image;

[0035] Step 3.6. The image after morphological operation is displayed as a discontinuous polygon. The discontinuous polygon is contour scanned, and curvature fitting is performed after contour scanning to obtain a closed boundary curve;

[0036] Step 3.7. Extract the image within the closed curve, use the grid to calculate and mark the area of ​​the region contained in the closed curve, and the area of ​​the region within the closed curve is the characteristic value of the plantar pressure characteristic region with the same boundary grayscale.

[0037] Furthermore, the establishment of the multivariate linear regression model in step 4 includes the following steps:

[0038] Step 4.1. Take the estimated height value as the dependent variable Y, and select the characteristic value of the plantar pressure characteristic area as the independent variable X, then the multivariate linear regression equation is as follows (4):

[0039] Y=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6+β7X7+ε……(4),

[0040] In the above formula (4), β0, β1, ..., β7 are regression parameters, X1 is the characteristic value of the first metatarsal region, X2 is the characteristic value of the heel region, X3 is the characteristic value of the second to third metatarsal region, X4 is the characteristic value of the first phalanx region, X5 is the characteristic value of the arch region, X6 is the characteristic value of the fourth to fifth metatarsal region, X7 is the characteristic value of the second to fifth phalanx region, ε is the random error of each measurement, and satisfies the Gauss-Markov hypothesis, that is, ε~N(0, σ 2 ), indicating that ε obeys a mathematical expectation of 0 and a variance of σ 2 Normal distribution of

[0041] Step 4.2. Collect n sets of independent experimental data and form a matrix equation group, as shown in the following equation (5):

[0042] Y=Xβ+ε……(5),

[0043] In the above formula (5),

[0044] Step 4.3. Use the least squares method to obtain regression parameters:

[0045] Step 4.3.1 The expression of the sum of squares of n observation errors Q is as follows (6):

[0046] Q=ε′ε=(Y-Xβ)′(Y-Xβ)……(6),

[0047] The smaller the Q value in the above formula (6), the smaller the error. The ideal state is Q = 0;

[0048] Step 4.3.2 Take the derivative of both sides of the above equation (6) with respect to β and obtain the following equation (7):

[0049]

[0050] Step 4.3.3: Further solve equation (7) to obtain the following equation (8):

[0051]

[0052] In the above formula (8), is the regression parameter matrix when Q takes the minimum value;

[0053] Step 4.3.4 From the above formula, we can obtain the empirical regression equation (9):

[0054]

[0055] In the above formula (9), Estimated experience value for height, is the empirical regression parameter;

[0056] Step 4.4. Perform a significance test on the empirical regression equation. If the significance level reaches the threshold, the empirical equation is determined to be a multiple linear regression model of the plantar pressure characteristic value and height.

[0057] Furthermore, in step 5, the neural network model is repeatedly used to obtain the plantar pressure characteristic value, and the height is estimated by the multiple linear regression model. Considering that the plantar pressure image found on the scene is a set of plantar pressure images of a plantar imprint, the processing of the plantar pressure image set needs to first screen, eliminate and standardize the obtained set of plantar pressure characteristic values, and then input them into the multiple linear regression model for height estimation, including the following steps:

[0058] Step 5.1. Perform the Romanovsky criterion (T test criterion) test on the obtained set of plantar pressure characteristic values. First, remove a suspicious value, and then test the removed value as an outlier according to the T distribution. Suppose the sample data is X i1 ,X i2 ,…,X in ,i∈{1,2,…,7}, if we think that X ij For suspicious values, calculate the average value of n-1 data The standard deviation S i，n-1 As shown in formula (10) and formula (11):

[0059]

[0060]

[0061] According to the T distribution test, the removed X ij Is it an outlier? Then X ij As an outlier, it is removed. Then X ij If the data are normal, they will be retained; where α is the significance level, n is the number of data, and q(n, α) is the detection coefficient; the T distribution test is used until all data are normal;

[0062] Step 5.2. Standardize the remaining data after removing the outliers, and calculate the root mean square of the remaining data. The formula for calculating the root mean square is as follows (12):

[0063]

[0064] In the above formula (12), m is the number of data remaining after removing outliers;

[0065] Step 5.3. Repeat steps 5.1. and 5.2. to obtain the respective variables X of the multiple linear regression model i , i∈{1, 2, …, 7}, and X i , the values ​​of i∈{1, 2, …, 7} are input into the multivariate linear regression model to obtain the estimated height.

[0066] Compared with the prior art, the height estimation method of the present invention has the following beneficial effects:

[0067] 1. The height estimation method of the present invention adopts the pressure characteristic value of the plantar pressure characteristic area to replace the traditional length characteristic value between plantar characteristic points, so that the acquisition of characteristic values ​​is more representative and scientific, and overcomes the problem that the traditional length characteristic value cannot accurately map the height change of the subject to be measured due to slight changes in physiological factors in a specific age range. At the same time, compared with the acquisition of length characteristic value, the acquisition of pressure characteristic value is more convenient and accurate. In view of the fact that no matter in what state the subject to be measured contacts the ground, surface contact will be generated, that is, a plantar pressure area will be generated, while the length characteristic value cannot be obtained in any state.

[0068] 2. The height estimation method described in the present invention adopts a neural network model with deep self-learning ability to analyze and judge the boundaries and areas of the plantar pressure characteristic area. With the continuous accumulation of data, the confidence of the judgment will become higher and higher, thereby effectively avoiding the human errors caused by the inconsistent judgment standards brought about by the traditional judgment of length characteristic values ​​being completed entirely by manual operation. In addition, the neural network model proposed in the present invention introduces the BN layer based on the VGG16 network, and the calculation efficiency is higher, the output results are more stable, and the requirements of work intensity on efficiency in a real-time environment can be met, providing technical support for saving a lot of manpower, material resources and time costs.

[0069] 3. The height estimation method described in the present invention fully considers the ability of the plantar pressure area to express pressure characteristics, creatively defines seven plantar pressure characteristic areas as linearly independent and non-correlated parameter variables for extracting plantar characteristic values, and integrates the plantar pressure areas with strong correlation into one area for consideration, thereby ensuring the accuracy, scientificity and convenience of sample collection in establishing a multivariate linear regression model. At the same time, a group of plantar pressure images will appear for the same plantar print. In the sample screening of the plantar pressure area characteristic values, the T distribution test and root mean square standardization are successively used to ensure the validity and accuracy of the sample data.

[0070] 4. The height estimation method described in the present invention proposes a specific method for image preprocessing, defines reference values ​​for screening plantar pressure images, and significantly improves the speed and accuracy of neural network interpretation by rotating and centering the plantar pressure images. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the height estimation method of the present invention;

[0072] Figure 2 A flowchart of collecting plantar pressure image samples and preprocessing them as described in step 1;

[0073] Figure 3 a is the VGG16 network model diagram described in step 2.1. Figure 3 b is the VGG16 network model diagram of the introduction of the BN layer in step 2.4;

[0074] Figure 4 It is a calculation flow chart of the neural network model described in step 3;

[0075] Figure 5 This is the regional distribution diagram of the plantar pressure characteristic values ​​described in step 4. DETAILED DESCRIPTION

[0076] The specific implementation of the height estimation method based on plantar pressure of the present invention is now described in detail with reference to the accompanying drawings of the specification.

[0077] like Figure 1 As shown, a height estimation method based on plantar pressure includes the following steps:

[0078] Step 1. Collect plantar pressure image samples and preprocess them;

[0079] Step 2. Build and train the neural network model;

[0080] Step 3. Input the preprocessed image samples into the trained neural network model to calculate the plantar pressure characteristic value;

[0081] Step 4. Establish a multiple linear regression model between plantar pressure characteristic values ​​and height;

[0082] Step 5. Continue to use the neural network model to calculate the plantar pressure characteristic value and estimate the height through the multivariate linear regression model.

[0083] Further, such as Figure 2 As shown, the step 1 of collecting plantar pressure image samples and preprocessing includes the following steps:

[0084] Step 1.1. Define the plantar pressure feature areas in the plantar pressure image sample of an adult in walking state, including: the first metatarsal area, the second to third metatarsal areas, the fourth to fifth metatarsal areas, the heel area, the first phalanx area, the second to fifth phalanx areas and the arch area;

[0085] Step 1.2. Screen the collected plantar pressure image samples according to the expression ability of the plantar pressure characteristic area on the pressure characteristics, and eliminate the plantar pressure images that lack the first metatarsal area, the first phalanx area or the heel area in the walking state;

[0086] Step 1.3. Remove the additional information such as the scale, watermark and grid line in the screened plantar pressure image that is not suitable for neural network interpretation;

[0087] Step 1.4. For the plantar pressure image whose footprint direction and footprint position are not in the center of the image, rotation correction and footprint centering operations are performed;

[0088] Step 1.5. Using the point 10 pixels outward from the most prominent point on the upper and lower sides of the footprint as the tangent point, crop the image into a rectangular image of uniform size and enhance the resolution.

[0089] Further, such as Figure 3 As shown, step 2 describes establishing and training a neural network model, including the following steps:

[0090] Step 2.1. Figure 3 As shown in figure a, the VGG16 network has a total of 16 layers, including 13 convolutional layers and 3 fully connected layers, of which the 13 convolutional layers are divided into 5 convolutional blocks. The first convolutional block contains 2 3*3*64 convolutional layers, the second convolutional block contains 2 3*3*128 convolutional layers, the third convolutional block contains 3 3*3*256 convolutional layers, the fourth convolutional block contains 3 3*3*256 convolutional layers, and the fifth convolutional block contains 3 3*3*256 convolutional layers. The first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, and the fifth convolutional block are all connected to a maximum pooling layer. The maximum pooling layer connected to the fifth convolutional block is connected to 3 fully connected layers. The last fully connected layer among the 3 fully connected layers is connected to the Softmax layer. The output of each fully connected layer is used as the input of the next fully connected layer.

[0091] Step 2.2. Input layer: Input a local context area image of size 224*224*3 to the VGG16 network. The input image is convolved with the convolution block to obtain multiple feature maps corresponding to each convolution block.

[0092] Step 2.3. The activation function performs nonlinear processing on multiple feature maps corresponding to each convolution block. The activation function uses the ReLU function. After the input image passes through the convolution and pooling layers, the feature map obtained by the last fully connected layer is further processed by the Softmax layer to obtain the classification result, which is output through the output layer.

[0093] Step 2.4. Figure 3 As shown in b, in view of the feature that the plantar pressure image contains few chromatic aberration pixels, based on the VGG16 network model, a batch normalization layer (BatchNormalization, BN) is introduced after each pooling layer to accelerate the training of the neural network model. The neural network model training input sample size is set to be no less than 4000, the batch size is 128 images as 1 Batch, the number of iterations is no less than 350 times, and the learning utilization rate is 0.0001.

[0094] Further, such as Figure 4As shown, the calculation of the plantar pressure characteristic value in step 3 includes the following steps:

[0095] Step 3.1. Perform denoising on the preprocessed image samples. First, use a bilateral mean filter to reduce the noise interference in the spatial domain of the image while maintaining the edge texture blur information of the plantar image. Then, use a median filter to eliminate the salt-and-pepper noise in the image while retaining the sharpness of the edge of the plantar image.

[0096] Step 3.2. Convert the preprocessed image samples into grayscale images and mark the plantar pressure feature area. The marked area contains the fuzzy texture information of the edge of the plantar pressure feature area and is tangent to the arc with the largest curvature of the edge of the plantar pressure feature area.

[0097] Step 3.3. Perform histogram equalization on the marked area, scan all pixels in the marked area, and calculate the normalized histogram H of the marked area. The calculation process is to first list the differential histogram H(r k ) expression, and then the differential histogram H(r k ) integral, the calculation process is expressed as follows (1) and (2):

[0098]

[0099] H′(i)=∑ 0≤j≤i H(j)……(2),

[0100] In the above formulas (1) and (2), n is the total number of images, k is the kth grayscale level in the image, and n k is the number of gray levels, i represents the highest gray level in the region, and j represents the gray level variable in the region;

[0101] Finally, the image is transformed, which is expressed as the following formula (3):

[0102] dst(x,y)=H′(src(x,y))……(3),

[0103] In the above formula (3), dst(x, y) is the output image, and src(x, y) represents the input image;

[0104] Step 3.4. Use the neighborhood block Gaussian weighted method to perform local adaptive binary threshold segmentation on the image of the marked area, divide the image of the marked area into the area of ​​interest and the remaining area, and extract the area of ​​interest for further analysis;

[0105] Step 3.5. Perform median filtering on the area of ​​interest again. After obtaining an interference-free image, perform 3×3 kernel erosion first, and then perform 12×12 kernel dilation to complete the morphological operation of the image;

[0106] Step 3.6. The image after morphological operation is displayed as a discontinuous polygon. The discontinuous polygon is contour scanned, and curvature fitting is performed after contour scanning to obtain a closed boundary curve;

[0107] Step 3.7. Extract the image within the closed curve, use the grid to calculate and mark the area of ​​the region contained in the closed curve, and the area of ​​the region within the closed curve is the characteristic value of the plantar pressure characteristic region with the same boundary grayscale.

[0108] Furthermore, the establishment of the multivariate linear regression model in step 4 includes the following steps:

[0109] Step 4.1. Take the estimated height value as the dependent variable Y, and select the characteristic value of the plantar pressure characteristic area as the independent variable X, then the multivariate linear regression equation is as follows (4):

[0110] Y=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6+β7X7+ε……(4),

[0111] In the above formula (4), β0, β1, ..., β7 are regression parameters, such as Figure 5 As shown in the figure, X1 is the characteristic value of the first metatarsal region, X2 is the characteristic value of the heel region, X3 is the characteristic value of the second to third metatarsal region, X4 is the characteristic value of the first phalanx region, X5 is the characteristic value of the arch region, X6 is the characteristic value of the fourth to fifth metatarsal region, X7 is the characteristic value of the second to fifth phalanx region, ε is the random error of each measurement, and satisfies the Gauss-Markov hypothesis, ε~N(0,σ 2 ), indicating that ε obeys a mathematical expectation of 0 and a variance of σ 2 Normal distribution of

[0112] Step 4.2. Collect n sets of independent experimental data and form a matrix equation group, as shown in the following equation (5):

[0113] Y=Xβ+ε……(5),

[0114] In the above formula (5),

[0115] Step 4.3. Use the least squares method to obtain regression parameters:

[0116] Step 4.3.1 The expression of the sum of squares of n observation errors Q is as follows (6):

[0117] Q=ε'ε=(Y-Xβ)'(Y-Xβ)......(6),

[0118] The smaller the Q value in the above formula (6), the smaller the error. The ideal state is Q = 0;

[0119] Step 4.3.2 Take the derivative of both sides of the above equation (6) with respect to β and obtain the following equation (7):

[0120]

[0121] Step 4.3.3: Further solve equation (7) to obtain the following equation (8):

[0122]

[0123] In the above formula (8), is the regression parameter matrix when Q takes the minimum value;

[0124] Step 4.3.4 Thus, we get the empirical regression equation (9):

[0125]

[0126] In the above formula (9), Estimated experience value for height, is the empirical regression parameter;

[0127] Step 4.4. Perform a significance test on the empirical regression equation. If the significance level reaches the threshold, the empirical equation is determined to be a multiple linear regression model of the plantar pressure characteristic value and height.

[0128] Furthermore, in step 5, the neural network model is repeatedly used to obtain the plantar pressure characteristic value, and the height is estimated by the multiple linear regression model. Considering that the plantar pressure image found on the scene is a set of plantar pressure images of a plantar imprint, the processing of the plantar pressure image set needs to first screen, eliminate and standardize the obtained set of plantar pressure characteristic values, and then input them into the multiple linear regression model for height estimation, including the following steps:

[0129] Step 5.1. Perform the Romanovsky criterion (T test criterion) test on the obtained set of plantar pressure characteristic values. First, remove a suspicious value, and then test the removed value as an outlier according to the T distribution. Suppose the sample data is X i1 ,X i2 ,…,X in, i∈{1,2,…,7}, if we think X ij For suspicious values, calculate the average value of n-1 data The standard deviation S i，n-1 As shown in the following formulas (10) and (11):

[0130]

[0131]

[0132] According to the T distribution test, the removed X ij Is it an outlier? Then X ij As an outlier, it is removed. Then X ij If the data are normal, they will be retained; where α is the significance level, n is the number of data, and q(n, α) is the detection coefficient; the T distribution test is used until all data are normal;

[0133] Step 5.2. Standardize the remaining data after removing the outliers by calculating the root mean square of the remaining data. The formula for calculating the root mean square is as follows (12):

[0134]

[0135] In the above formula (12), m is the number of data remaining after removing outliers;

[0136] Step 5.3. Repeat steps 5.1. and 5.2. to obtain the respective variables X of the multiple linear regression model i , i∈{1, 2, …, 7}, and X i , the values ​​of i∈{1, 2, …, 7} are input into the multivariate linear regression model to obtain the estimated height.

[0137] The present invention is not limited to the above-mentioned embodiments. Without departing from the essential content of the present invention, any deformation, improvement and substitution that can be thought of by those skilled in the art shall fall into the protection scope of the present invention.

Claims

1. A height estimation method based on plantar pressure, characterized in that: The height estimation method comprises the following steps: Step 1. Collect plantar pressure image samples and preprocess them: Step 1.

1. Define the plantar pressure feature areas in the plantar pressure image sample of an adult in walking state, including: the first metatarsal area, the second to third metatarsal areas, the fourth to fifth metatarsal areas, the heel area, the first phalanx area, the second to fifth phalanx areas and the arch area; Step 1.

2. Screen the collected plantar pressure image samples according to the expression ability of the plantar pressure characteristic area on the pressure characteristics, and eliminate the plantar pressure images that lack the first metatarsal area, the first phalanx area or the heel area in the walking state; Step 1.

3. Remove the additional information such as the scale, watermark and grid line in the screened plantar pressure image that is not suitable for neural network interpretation; Step 1.

4. For the plantar pressure image whose footprint direction and footprint position are not in the center of the image, rotation correction and footprint centering operations are performed; Step 1.

5. Cut the image into a rectangular image of uniform size, taking the most prominent point 10 pixels outward from the top and bottom of the footprint as the tangent point, and enhance the resolution; Step 2. Build and train the neural network model: Step 2.

1. The VGG16 network has a total of 16 layers, including 13 convolutional layers and 3 fully connected layers. The 13 convolutional layers are divided into 5 convolutional blocks. The first convolutional block contains 2 3*3*64 convolutional layers, the second convolutional block contains 2 3*3*128 convolutional layers, the third convolutional block contains 3 3*3*256 convolutional layers, the fourth convolutional block contains 3 3*3*256 convolutional layers, and the fifth convolutional block contains 3 3*3*256 convolutional layers. The first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, and the fifth convolutional block are all connected to a maximum pooling layer. The maximum pooling layer connected to the fifth convolutional block is connected to 3 fully connected layers. The last fully connected layer among the 3 fully connected layers is connected to the Softmax layer. The output of each fully connected layer is used as the input of the next fully connected layer. Step 2.

2. Input layer: Input a 224*224*3 local context area image to the VGG16 network. The input image is convolved with the convolution block to obtain multiple feature maps corresponding to each convolution block. Step 2.

3. The activation function performs nonlinear processing on multiple feature maps corresponding to each convolution block. The activation function uses the ReLU function. After the input image passes through the convolution and pooling layers, the feature map obtained by the last fully connected layer is further processed by the Softmax layer to obtain the classification result, which is output through the output layer. Step 2.

4. In view of the fact that the plantar pressure image contains few color difference pixels, based on the VGG16 network model, a batch normalization layer is introduced after each pooling layer to accelerate the training of the neural network model. The neural network model training input sample size is set to be no less than 4000, the batch size is 128 images per batch, the number of iterations is no less than 350, and the learning utilization rate is 0.0001; Step 3. Input the preprocessed image samples into the neural network model trained in step 2.4 to calculate the plantar pressure characteristic value: Step 3.

1. Perform denoising on the preprocessed image samples. First, use a bilateral mean filter to reduce the noise interference in the spatial domain of the image while maintaining the edge texture blur information of the plantar image. Then, use a median filter to eliminate the salt and pepper noise in the image while retaining the sharpness of the edge of the plantar image. Step 3.

2. Convert the preprocessed image samples into grayscale images and mark the plantar pressure feature area. The marked area contains the fuzzy texture information of the edge of the plantar pressure feature area and is tangent to the arc with the largest curvature of the edge of the plantar pressure feature area. Step 3.

3. Perform histogram equalization on the marked area, scan all pixels in the marked area, and calculate the normalized histogram H of the marked area. The calculation process is to first list the differential histogram H(r k ) expression, and then the differential histogram H(r k ) integral, the calculation process is expressed as follows (1) and (2): H′(i)=∑ 0≤j≤i H(j)……(2), In the above formulas (1) and (2), n is the total number of images, k is the kth gray level in the image, nk is the number of gray levels, i represents the highest gray level in the region, and j represents the gray level variable in the region; Finally, the image is transformed, which is expressed as the following formula (3): dst(x,y)=H′(src(x,y))……(3), In the above formula (3), dst(x,y) is the output image, and src(x,y) represents the input image; Step 3.

4. Use the neighborhood block Gaussian weighted method to perform local adaptive binary threshold segmentation on the image of the marked area, divide the image of the marked area into the focus area and the remaining area, and extract the focus area for further analysis; Step 3.

5. Perform median filtering on the area of ​​interest again. After obtaining an interference-free image, perform 3×3 kernel erosion first, and then perform 12×12 kernel dilation to complete the morphological operation of the image; Step 3.

6. The image after morphological operation is displayed as a discontinuous polygon. The discontinuous polygon is contour scanned, and curvature fitting is performed after contour scanning to obtain a closed boundary curve; Step 3.

7. Extract the image within the closed curve, use the grid to calculate and mark the area of ​​the region contained in the closed curve, and the area of ​​the region within the closed curve is the characteristic value of the plantar pressure characteristic region with the same boundary grayscale; Step 4. Establish a multiple linear regression model between the characteristic values ​​of the plantar pressure characteristic area and height: Step 4.

1. Take the estimated height value as the dependent variable Y, and select the characteristic value of the plantar pressure characteristic area as the independent variable X, then the multivariate linear regression equation is as follows (4): Y=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6+β7X7+ε……(4), In the above formula (4), β0, β1, ..., β7 are regression parameters, X1 is the characteristic value of the first metatarsal region, X2 is the characteristic value of the heel region, X3 is the characteristic value of the second to third metatarsal region, X4 is the characteristic value of the first phalanx region, X5 is the characteristic value of the arch region, X6 is the characteristic value of the fourth to fifth metatarsal region, X7 is the characteristic value of the second to fifth phalanx region, ε is the random error of each measurement, and satisfies the Gauss-Markov hypothesis, that is, ε~N(0,σ 2 ), indicating that ε obeys a mathematical expectation of 0 and a variance of σ 2 Normal distribution of Step 4.

2. Collect n sets of independent experimental data and form a matrix equation group, as shown in the following equation (5): Y=Xβ+ε……(5), In the above formula (5), Step 4.

3. Use the least squares method to obtain regression parameters: Step 4.3.1 The expression of the sum of squares of n observation errors Q is as follows (6): Q=ε'ε=(Y-Xβ)'(Y-Xβ)......(6), The smaller the Q value in the above formula (6), the smaller the error. The ideal state is Q = 0; Step 4.3.2 Take the derivative of both sides of the above equation (6) with respect to β and obtain the following equation (7): Step 4.3.3: Further solve equation (7) to obtain the following equation (8): In the above formula (8), is the regression parameter matrix when Q takes the minimum value; Step 4.3.4 From the above formula, we can obtain the empirical regression equation (9): In the above formula (9), Estimated experience value for height, is the empirical regression parameter; Step 4.

4. Perform a significance test on the empirical regression equation. If the significance level reaches the threshold, the empirical equation is determined to be a multiple linear regression model of plantar pressure characteristic value and height; Step 5. Continue to use the neural network model to calculate the plantar pressure characteristic value and estimate the height through the multivariate linear regression model: Step 5.

1. Perform the Romanovsky criterion test on the obtained set of plantar pressure characteristic values. First, remove a suspicious value, and then test the removed value as an outlier according to the T distribution. Suppose the sample data is X i1 ,X i2 ,…,X in ,i∈{1,2,…,7}, if we think that X ij For suspicious values, calculate the average value of n-1 data The standard deviation S i,n-1 As shown in formula (10) and formula (11): According to the T distribution test, the removed X ij Is it an outlier? Then X ij As an outlier, it is removed. Then X ij If the data are normal, they will be retained; where α is the significance level, n is the number of data, and q(n,α) is the detection coefficient; the T distribution test is used until all data are normal; Step 5.

2. Standardize the remaining data after removing the outliers, and calculate the root mean square of the remaining data. The formula for calculating the root mean square is as follows (12): In the above formula (12), m is the number of data remaining after removing outliers; Step 5.

3. Repeat steps 5.

1. and 5.

2. to obtain the respective variables X of the multiple linear regression model i , i∈{1,2,…,7}, and X i , the value of i∈{1,2,…,7} is input into the multivariate linear regression model to obtain the estimated height.

Images