CN109620244B - A Conditional Generative Adversarial Network and SVM-Based Approach for Infant Abnormal Behavior Detection - Google Patents
A Conditional Generative Adversarial Network and SVM-Based Approach for Infant Abnormal Behavior Detection Download PDFInfo
- Publication number
- CN109620244B CN109620244B CN201811494749.0A CN201811494749A CN109620244B CN 109620244 B CN109620244 B CN 109620244B CN 201811494749 A CN201811494749 A CN 201811494749A CN 109620244 B CN109620244 B CN 109620244B
- Authority
- CN
- China
- Prior art keywords
- abnormal
- whole body
- wavelet
- normal
- limbs
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Measuring devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor or mobility of a limb
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Measuring devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor or mobility of a limb
- A61B5/1113—Local tracking of patients, e.g. in a hospital or private home
- A61B5/1114—Tracking parts of the body
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/103—Measuring devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
- A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor or mobility of a limb
- A61B5/1126—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor or mobility of a limb using a particular sensing technique
- A61B5/1127—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor or mobility of a limb using a particular sensing technique using markers
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Veterinary Medicine (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Physiology (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Public Health (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Dentistry (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Fuzzy Systems (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Signal Processing (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
- Processing Or Creating Images (AREA)
Abstract
基于条件生成对抗网络和SVM的婴儿异常行为检测方法属视频图像处理和深度学习技术领域，本发明通过对视频中婴儿运动轨迹进行分析判断婴儿行为是否异常，首先获取婴儿视频，进行合理长度的截取并转化为帧图像，对四肢和全身进行标记建立样本库；然后利用条件生成对抗网络对婴儿的四肢和全身进行目标跟踪；再对得到的目标运动轨迹进行小波近似波形和小波功率谱计算，将得到的特征用支持向量机SVM进行分类，综合判断；本发明对婴儿四肢和全身信息进行运动轨迹检测，比单一肢体检测信息更全面，小波域和功率谱域结合训练，使检测精确度提高，检测婴儿行为是否异常，及早进行干预，对预防婴儿脑瘫等疾病有很重要的意义。
The abnormal infant behavior detection method based on conditional generative adversarial network and SVM belongs to the technical field of video image processing and deep learning. The present invention judges whether the infant behavior is abnormal by analyzing the motion trajectory of the infant in the video. First, the infant video is obtained and intercepted with a reasonable length. And convert it into frame images, mark the limbs and the whole body to establish a sample library; then use the conditional generative adversarial network to track the baby's limbs and whole body; then calculate the wavelet approximate waveform and wavelet power spectrum of the obtained target motion trajectory, The obtained features are classified and judged comprehensively by the support vector machine SVM; the present invention performs motion trajectory detection on infant limbs and whole body information, which is more comprehensive than single limb detection information, and the combination of wavelet domain and power spectrum domain training improves detection accuracy, It is of great significance to detect whether the infant's behavior is abnormal and to intervene as soon as possible to prevent diseases such as infant cerebral palsy.
Description
Technical Field
The invention belongs to the technical field of video image processing and deep learning, and particularly relates to a method for detecting abnormal behaviors of infants based on a condition generation countermeasure network and an SVM.
Background
The abnormal behavior of the infant mainly refers to that small-amplitude middle-speed movement in various directions with variable acceleration is not seen in the movement of the infant, the movement of the infant is not suitable for the age of the infant, other movement forms (such as limb midline movement, hand and knee contact, visual search, finger scratching and clothes grabbing and the like) are not available, and the whole movement is poor in fluency. The abnormal behavior of the infant is damaged corresponding to the brain, severe patients can cause cerebral palsy, and cerebral palsy diseases can be diagnosed only after the child ages from one to two, so that the early detection of the abnormal behavior of the infant and timely intervention and treatment have strong practical significance.
In response to this problem, researchers have proposed several methods for detecting abnormal behavior in infants, which can be broadly divided into three types: whole body motion quality assessment, wearable sensor assessment, combined with pattern recognition assessment. The first method is to adopt a specific video recording mode for the baby, and judge whether the behavior is abnormal or not by using the whole body movement quality evaluation criterion for the video recording result, and the mode mainly depends on observation and has certain subjectivity. The second method is to wear a sensor device on the infant to observe the parameters, but the wearable method itself interferes with the movement of the infant to some extent, causing inaccuracy in the prediction result. The third method is to extract the motion characteristics of the baby by using a computer to perform pattern recognition analysis, which does not interfere with the movement of the baby and has objectivity, but in the process of extracting the motion characteristics and recognizing, only a limited number of body parts are observed, the analysis of the whole body motion is not performed, and certain specificity is achieved.
Due to the defects of the above algorithm, it is difficult to achieve the desired effect in practical applications, so improvement is necessary.
Disclosure of Invention
The invention aims to provide a baby abnormal behavior detection method based on a condition generation countermeasure network CGAN, which is combined with a supervised SVM classification method to improve the accuracy of baby abnormal behavior detection.
A baby abnormal behavior detection method based on condition generation countermeasure network and SVM is characterized in that a training sample library required by target tracking is constructed in advance, the condition generation countermeasure network is utilized to track the four limbs and the whole body of a baby, the training sample library comprises the limbs and the whole body marked by the baby, motion trail information is extracted by utilizing wavelet approximate waveform and wavelet power spectrum analysis, and the characteristics of the motion trail information are classified by a support vector machine SVM, and the method comprises the following steps:
1.1, acquiring a baby video and carrying out unified preprocessing;
1.2, intercepting the baby video obtained in the step 1.1 for 15s, uniformly naming, and uniformly naming the images converted into frames;
1.3 tracking of baby motion trail: for the frame image obtained in the step 1.2, a confrontation network CGAN is generated by using conditions to track the four limbs and the whole body movement locus of the baby respectively, and the method specifically comprises the following steps:
1.3.1 constructing a training sample library required by target tracking: marking the left hand, the right hand, the left leg, the right leg and the whole body of the baby in the frame image obtained in the step 1.2, forming a training sample library by the marked limbs and the whole body of the baby as a target data set input to CGAN, and using the corresponding label as a condition Y;
1.3.2 generative model design: randomly dividing each frame of image containing the baby to be used as a pseudo target data set, and inputting the pseudo target data set and the condition Y into a discrimination model device through a convolution layer;
1.3.3 discriminant model design: sending the target data set and the condition Y into a discrimination model to judge limbs and the whole body, sending the pseudo target data set into a discriminator, and judging whether the target is a target or not;
1.3.4 judging whether the target is the target or not, calculating errors to enable the errors to accord with a formula, wherein the specific formula is as follows:
optimizing D:
maxDV(D，G)＝Ex～pdata(x)[log(D(x))]+Ez～pz(z)[log(1-D(G(z)))]
optimizing G:
minGV(D，G)＝Ez～pz(z)[log(1-D(G(z)))]
wherein: v (D, G) represents a loss function; pdata (x) is the true sample distribution; pz (z) is the pseudo-sample distribution; d (x) represents the real sample data in the discriminator; d (g (z)) represents pseudo sample data in the discriminator; e represents expectation;
performing model parameter adjustment according to the optimization conditions, wherein parameters of the generated model G and the discrimination model D are shared;
1.3.5 if the error is too large, feeding back the error to the input of the generation model, reconstructing a pseudo target data set, judging again until the positions of the four limbs and the whole body of the baby in the pseudo target data set are found, and recording the positions and the motion tracks of the left hand, the right hand, the left leg, the right leg and the whole body of the baby in each frame;
1.4, analyzing the motion trail information: the method specifically comprises the following steps of storing the position information of continuous y-axis coordinate change during movement of the four limbs and the whole body of the baby tracked in the step 1.3, and calculating the wavelet approximation waveform and the wavelet power spectrum of a continuous change waveform diagram formed by the position information of the y-axis coordinate, wherein the continuous change waveform diagram comprises the following steps:
1.4.1 because the coordinate change of the x axis is not obvious, only selecting the coordinate change diagram of the y axis for analysis, firstly, analyzing the approximate waveform of the wavelet, and analyzing the tracked waveform by using harr wavelet to obtain the approximate waveform of the wavelet;
1.4.2 for the y-axis coordinate change graphs of limbs and the whole body, power spectrum information is obtained by utilizing a power spectrogram based on wavelets;
1.5 extracting characteristic vectors from the obtained wavelet approximate oscillogram and wavelet power spectrogram, and training and learning by using a Support Vector Machine (SVM), specifically comprising the following steps:
1.5.1, dividing the sample into normal and abnormal samples for marking, and setting a normal sample label as 1 and an abnormal sample label as-1;
1.5.2 dividing the sample into a training set and a testing set, normalizing the data, and obtaining the highest accuracy by adjusting the values of parameters c and g in the SVM, thereby obtaining the optimal training model;
1.6 comprehensive judgment of abnormal behaviors of infants: according to the optimal training model obtained in the step 1.5.2, different weights are set for different accuracies, and weighting judgment is carried out, and the method specifically comprises the following steps:
1.6.1 for the SVM model trained from the wavelet approximation waveform obtained in step 1.4.1, different weight coefficients are set according to different accuracies of limbs and the whole body, specifically: left upper limb a 1: 0.35; right upper limb a 2: 0.01; left lower limb a 3: 0.2; right lower limb a 4: 0.35; whole body a 5: 0.09; the judgment result vectors of the four limbs and the whole body are respectively represented by Y1 to Y5, and the calculation formula is as follows:
Y1＝(test label+predict label)/2
wherein: test label is the actual label of the test sample; the prediction label is a label predicted by the test sample; y2 through Y5 were calculated in the same manner as above;
the five resulting vectors are weighted, as follows:
Y＝0.35*Y1+0.01*Y2+0.2*Y3+0.35*Y4+0.09*Y5
wherein: the multiplication operation is represented, Y is a judgment value predicted by the wavelet details, a judgment standard is defined, if-1 < Y < -0.3, the baby behavior is judged to be in an abnormal state, if 0.3< Y <1, the baby behavior is judged to be in a normal state, and the rest are all regarded as judgment error states;
1.6.2 for the SVM model trained from the wavelet power spectrum obtained in the step 1.4.2, different weight coefficients are set according to different accuracy of limbs and the whole body, specifically: left upper limb P1: 0.35; right upper limb P2: 0.01; left lower limb P3: 0.35; right lower limb P4: 0.2; whole body P5: 0.09; the judgment result vectors of the four limbs and the whole body are respectively expressed by X1 to X5, and the calculation formula is as follows:
X1＝(test label+predict label)/2
wherein: test label is the actual label of the test sample; the prediction label is a label predicted by the test sample; x2 through X5 were calculated in the same manner as above;
the five resulting vectors are weighted, as follows:
X＝0.35*X1+0.01*X2+0.35*X3+0.2*X4+0.09*X5
wherein: expressing multiplication, wherein X is a judgment value predicted by the wavelet power spectrum, a judgment standard is specified, if-1 is more than X and less than-0.3, the baby behavior is judged to be in an abnormal state, if 0.3 is more than X and less than 1, the baby behavior is judged to be in a normal state, and the rest are all considered to be in a judgment error state;
and comprehensively judging the X and the Y, and if the test sample at least meets one condition of the X and the Y, determining that the judgment result is correct, and distinguishing whether the behavior of the baby is normal.
The generative model design and discriminative model design of step 1.3.2 and step 1.3.3 specifically includes the following steps:
2.1 generative model design: wherein 6 layers of convolution layers are arranged, and the step length is set as 1; 6 layers of pooling layers, the size of the pooling window being 2 x 2; the network applies a corrected Linear unit relu (corrected Linear unit) activation function, which can obtain good results and faster convergence speed, and the specific operation formula is as follows:
F(Z)＝σ(W*Z+b)
wherein: w is the convolution kernel; is a convolution operation; z is a feature vector; b is an offset; σ is a ReLU activation function;
2.2 discriminant model design: wherein 5 layers of convolution layers are arranged, and the step length is set as 1; 5 layers of pooling layers, the size of the pooling window being 2 x 2; applied in the network is a correcting Linear unit ReLU (rectified Linear Unit) activation function.
Step 1.4 the specific calculation of the wavelet approximation waveform map and the wavelet power spectrogram comprises the following steps:
3.1 analyzing the tracked oscillogram by harr wavelets, constructing a five-layer pyramid according to a Mallat pyramid decomposition algorithm of discrete wavelet transformation, extracting a wavelet approximation signal of a fifth layer, corresponding to four limbs and the whole body, and respectively recording as: abnormal left upper limb: a01; abnormal right upper limb: a02; abnormal left lower limb: a03; abnormal right lower limb: a04; abnormal whole body: a05; normal left upper limb: a11; normal right upper limb: a12; normal left lower limb: a13; normal right lower limb: a14; normal whole body: a15;
3.2 for the y-axis coordinate change graphs of limbs and the whole body, utilizing a wavelet-based power spectrogram, wherein the set sampling length is the video total frame length 375, the sampling frequency is 1000, the sampling interval is 1/1000, and the obtained power spectrograms are respectively recorded as: abnormal left upper limb: p01; abnormal right upper limb: p02; abnormal left lower limb: p03; abnormal right lower limb: p04; abnormal whole body: p05; normal left upper limb: p11; normal right upper limb: p12; normal left lower limb: p13; normal right lower limb: p14; normal whole body: p15.
For generative model design and discriminative model design, the method comprises the following steps:
step A1, generating model design: wherein 6 layers of convolution layers are arranged, and the step length is set as 1; 6 layers of pooling layers, the size of the pooling window being 2 x 2; the network applies a corrected Linear unit relu (corrected Linear unit) activation function, which can obtain good results and faster convergence speed, and the specific operation formula is as follows:
F(Z)＝σ(W*Z+b)
wherein: w is the convolution kernel; is a convolution operation; z is a feature vector; b is an offset; σ is a ReLU activation function;
step A2, design of a discriminant model: wherein 5 layers of convolution layers are arranged, and the step length is set as 1; 5 layers of pooling layers, the size of the pooling window being 2 x 2; applied in the network is a correcting Linear unit ReLU (rectified Linear Unit) activation function.
Further, the specific calculation of the wavelet approximate oscillogram and the wavelet power spectrogram comprises the following steps:
step R1, analyzing the tracked oscillogram by harr wavelets, constructing a five-layer pyramid according to a Mallat pyramid decomposition algorithm of discrete wavelet transformation, extracting a fifth-layer wavelet approximate waveform corresponding to limbs and the whole body, and respectively recording as: abnormal left upper limb: a01; abnormal right upper limb: a02; abnormal left lower limb: a03; abnormal right lower limb: a04; abnormal whole body: a05; normal left upper limb: a11; normal right upper limb: a12; normal left lower limb: a13; normal right lower limb: a14; normal whole body: a15;
step B2, for the y-axis coordinate change diagrams of limbs and the whole body, utilizing a wavelet-based power spectrogram, wherein the set sampling length is the video total frame length 375, the sampling frequency is 1000, the sampling interval is 1/1000, and the obtained power spectrograms are respectively marked as: abnormal left upper limb: p01; abnormal right upper limb: p02; abnormal left lower limb: p03; abnormal right lower limb: p04; abnormal whole body: p05; normal left upper limb: p11; normal right upper limb: p12; normal left lower limb: p13; normal right lower limb: p14; normal whole body: p15.
The invention adopts a baby abnormal behavior detection method based on condition generation countermeasure network and SVM, firstly preprocessing the acquired baby video, then, a conditional generation countermeasure network CGAN is used for respectively tracking the target movement locus of the four limbs and the whole body of the baby in the video, the obtained movement locus information is stored, then the movement locus information is extracted by utilizing wavelet transformation, establishing a sample set for the extracted wavelet approximate waveform, training the sample set by using a set SVM (support vector machine), solving a power spectrum for motion trajectory information by using the wavelet to obtain a characteristic establishing sample set, training the sample set by using the set SVM support vector machine, testing two trained models, according to the difference of the accuracy of the two models, different weight parameters are set for weighting judgment, so that the optimal training result is obtained;
the invention combines the four limbs and the whole body information of the baby to detect the movement track, the information obtained by the detection is more comprehensive than the information obtained by single limb detection, the track tracking is more accurate by using CGAN network semi-supervised learning, and the characteristics have more abstract specificity by combining a wavelet domain and a power spectrum domain, meanwhile, an SVM support vector machine is used to classify the characteristics, the detection result is weighted and judged, the false detection rate is reduced, whether the baby behavior is abnormal or not is detected, the intervention is carried out as soon as possible, and the invention has important significance for preventing the diseases such as the cerebral palsy of the baby.
Drawings
FIG. 1 is a flow chart of a method for detecting abnormal behavior of an infant based on a condition-generated countermeasure network and SVM
FIG. 2 is a flow chart for tracking a motion trajectory using CGAN
FIG. 3 is an image of single frame infant left upper limb tracking
FIG. 4 is a y-axis motion trace image of a detected baby
FIG. 5 is a diagram of approximate wavelet waveform obtained by wavelet transform
FIG. 6 is a schematic diagram of wavelet power spectrum
FIG. 7 is a flowchart for determining whether baby behavior is abnormal
Detailed Description
The following describes the implementation process of the present invention with reference to the attached drawings.
An infant abnormal behavior detection method based on condition generation countermeasure network and SVM, which integrally realizes a flow, as shown in FIG. 1, and comprises the following steps:
1. and acquiring a baby video and carrying out unified preprocessing.
2. And (3) intercepting the baby video in the step (1) for 15s, uniformly naming, and uniformly naming the images converted into frames.
3. Tracking the motion trail of the baby: for the frame image obtained in step 2, a confrontation network CGAN is generated by using conditions to track the four limbs and the whole body movement locus of the infant, and a flow chart is shown in fig. 2, and specifically includes the following steps:
3.1 constructing a training sample library required by target tracking, marking the left hand, the right hand, the left leg, the right leg and the whole body of the baby in the frame image obtained in the step 2, forming the training sample library by the marked limbs and the whole body of the baby as a target data set input into the CGAN, and taking the corresponding label as a condition Y;
3.2 generative model design: randomly dividing each frame of image containing the baby to be used as a pseudo target data set, and inputting the pseudo target data set and the condition Y into a discrimination model device through a convolution layer;
3.3 discriminant model design: sending the target data set and the condition Y into a discrimination model to judge limbs and the whole body, sending the pseudo target data set into a discriminator, and judging whether the target is a target or not;
3.4 judging whether the target is the target or not, calculating the error to enable the error to accord with a formula, wherein the specific formula is as follows:
optimizing D:
maxDV(D，G)＝Ex～pdata(x)[log(D(x))]+Ez～pz(z)[log(1-D(G(z)))]
optimizing G:
minGV(D，G)＝Ez～pz(z)[log(1-D(G(z)))]
wherein: v (D, G) represents a loss function; pdata (x) is the true sample distribution; pz (z) is the pseudo-sample distribution; d (x) represents the real sample data in the discriminator; d (g (z)) represents pseudo sample data in the discriminator; e represents expectation.
The purpose of the formula is to minimize the error of the generated model to make the generated false target as true as possible, i.e. to find the target position as possible and to maximize the error of the discriminant model.
Performing model parameter adjustment according to the optimization conditions, wherein parameters of the generated model G and the discrimination model D are shared;
3.5 if the error is too large, feeding back the error to the input of the generation model, reconstructing the pseudo target data set, judging again until the positions of the limbs and the whole body of the baby in the pseudo target data set are found, as shown in fig. 3, recording the positions and the motion tracks of the left hand, the right hand, the left leg, the right leg and the whole body of the baby in each frame, wherein the image is a single frame of image tracked by the left upper limb of the baby.
4. Analyzing the motion track information: storing the position information of continuous y-axis coordinate change during movement (as shown in fig. 4) of the four limbs and the whole body of the baby tracked in the step 3, and calculating the wavelet approximation waveform and the wavelet power spectrum of a continuous change waveform diagram formed by the position information of the y-axis coordinate, specifically comprising the following steps:
4.1 because the coordinate change of the x axis is not obvious, only selecting the coordinate change diagram of the y axis for analysis, firstly, carrying out wavelet analysis, and analyzing the tracked oscillogram by using harr wavelets to obtain a wavelet approximate oscillogram;
and 4.2, solving power spectrum information by using a wavelet-based power spectrogram according to the y-axis coordinate change graphs of the limbs and the whole body.
5. Extracting feature vectors from the obtained wavelet approximate oscillogram and wavelet power spectrogram, and training and learning by using a Support Vector Machine (SVM), wherein the method specifically comprises the following steps:
5.1, dividing the sample into normal and abnormal samples for marking, setting a normal sample label as 1, and setting an abnormal sample label as-1;
5.2, dividing the sample into a training set and a testing set, normalizing the data, and obtaining the highest accuracy by adjusting the values of parameters c and g in the SVM, so as to obtain the optimal training model;
6. and (3) comprehensive judgment of abnormal behaviors of the infant: according to the optimal training model obtained in the step 5.2, different weights are set for different accuracies, and weighting judgment is carried out, and the method specifically comprises the following steps:
6.1 for the SVM model trained by the wavelet approximate waveform obtained in the step 4.1, different weight coefficients are set according to different accuracies of limbs and the whole body, specifically: left upper limb a 1: 0.35; right upper limb a 2: 0.01; left lower limb a 3: 0.2; right lower limb a 4: 0.35; whole body a 5: 0.09; the judgment result vectors of the four limbs and the whole body are respectively represented by Y1 to Y5, and the calculation formula is as follows:
Y1＝(test label+predict label)/2
wherein: test label is the actual label of the test sample; the prediction label is a label predicted by the test sample; y2 through Y5 were calculated in the same manner as above;
the five resulting vectors are weighted, as follows:
Y＝0.35*Y1+0.01*Y2+0.2*Y3+0.35*Y4+0.09*Y5
wherein: and (4) multiplication operation is performed, Y is a judgment value predicted by the wavelet details, a judgment standard is defined, if-1 < Y < -0.3, the baby behavior is judged to be in an abnormal state, if 0.3< Y <1, the baby behavior is judged to be in a normal state, and the rest are all considered to be in a judgment error state.
6.2 for the SVM model trained from the wavelet power spectrum obtained in the step 4.2, different weight coefficients are set according to different accuracy of limbs and the whole body, specifically: left upper limb P1: 0.35; right upper limb P2: 0.01; left lower limb P3: 0.35; right lower limb P4: 0.2; whole body P5: 0.09; the judgment result vectors of the four limbs and the whole body are respectively expressed by X1 to X5, and the calculation formula is as follows:
X1＝(test label+predict label)/2
wherein: test label is the actual label of the test sample; the prediction label is a label predicted by the test sample; the calculation of X2 through X5 is the same as above.
The five resulting vectors are weighted, as follows:
X＝0.35*X1+0.01*X2+0.35*X3+0.2*X4+0.09*X5
wherein: and (4) multiplication operation is performed, X is a judgment value predicted by the wavelet power spectrum, a judgment standard is defined, if-1 < X < -0.3, the baby behavior is judged to be in an abnormal state, if 0.3< X <1, the baby behavior is judged to be in a normal state, and the rest are all considered to be in a judgment error state.
And comprehensively judging X and Y, wherein a specific flow chart is shown in FIG. 7, if the test sample at least meets one condition of X and Y, the judgment result is considered to be correct, and whether the behavior of the baby is normal can be distinguished.
The invention relates to a generative model design and a discriminant model design, which comprises the following steps:
step A1, generating model design: wherein 6 layers of convolution layers are arranged, and the step length is set as 1; 6 layers of pooling layers, the size of the pooling window being 2 x 2; the network applies a corrected Linear unit relu (corrected Linear unit) activation function, which can obtain good results and faster convergence speed, and the specific operation formula is as follows:
F(Z)＝σ(W*Z+b)
wherein: w is the convolution kernel; is a convolution operation; z is a feature vector; b is an offset; σ is a ReLU activation function;
step A2, design of a discriminant model: wherein 5 layers of convolution layers are arranged, and the step length is set as 1; 5 layers of pooling layers, the size of the pooling window being 2 x 2; applied in the network is a correcting Linear unit ReLU (rectified Linear Unit) activation function.
The specific calculation of the wavelet approximate oscillogram and the wavelet power spectrogram in the invention comprises the following steps:
step B1, analyzing the tracked oscillogram by using hart wavelets, constructing a five-layer pyramid according to the Mallat pyramid decomposition algorithm of discrete wavelet transform, extracting a fifth-layer wavelet approximate waveform (as shown in fig. 5) corresponding to four limbs and the whole body, and respectively recording as: abnormal left upper limb: a01; abnormal right upper limb: a02; abnormal left lower limb: a03; abnormal right lower limb: a04; abnormal whole body: a05; normal left upper limb: a11; normal right upper limb: a12; normal left lower limb: a13; normal right lower limb: a14; normal whole body: A15.
step B2, for the y-axis coordinate change diagrams of limbs and the whole body, using a wavelet-based power spectrogram, wherein the set sampling length is the video total frame length 375, the sampling frequency is 1000, and the sampling interval is 1/1000, and the obtained power spectrograms (as shown in fig. 6) are respectively recorded as: abnormal left upper limb: p01; abnormal right upper limb: p02; abnormal left lower limb: p03; abnormal right lower limb: p04; abnormal whole body: p05; normal left upper limb: p11; normal right upper limb: p12; normal left lower limb: p13; normal right lower limb: p14; normal whole body: p15.
Claims (3)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811494749.0A CN109620244B (en) | 2018-12-07 | 2018-12-07 | A Conditional Generative Adversarial Network and SVM-Based Approach for Infant Abnormal Behavior Detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811494749.0A CN109620244B (en) | 2018-12-07 | 2018-12-07 | A Conditional Generative Adversarial Network and SVM-Based Approach for Infant Abnormal Behavior Detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109620244A CN109620244A (en) | 2019-04-16 |
| CN109620244B true CN109620244B (en) | 2021-07-30 |
Family
ID=66071933
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811494749.0A Expired - Fee Related CN109620244B (en) | 2018-12-07 | 2018-12-07 | A Conditional Generative Adversarial Network and SVM-Based Approach for Infant Abnormal Behavior Detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109620244B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109615019B (en) * | 2018-12-25 | 2022-05-31 | 吉林大学 | Abnormal behavior detection method based on space-time automatic encoder |
| CN110236558A (en) * | 2019-04-26 | 2019-09-17 | 平安科技（深圳）有限公司 | Method, device, storage medium and electronic equipment for predicting infant development |
| CN110414306B (en) * | 2019-04-26 | 2022-07-19 | 吉林大学 | Baby abnormal behavior detection method based on meanshift algorithm and SVM |
| CN111985269B (en) * | 2019-05-21 | 2024-12-03 | 顺丰科技有限公司 | Detection model construction method, detection method, device, server and medium |
| CN110277170A (en) * | 2019-06-28 | 2019-09-24 | 天津奥漫优悦科技有限公司 | A kind of all-around exercises assessment system based on AI artificial intelligence |
| US11596807B2 (en) * | 2019-11-25 | 2023-03-07 | Accuray Incorporated | Partial deformation maps for reconstructing motion-affected treatment dose |
| CN111353995B (en) * | 2020-03-31 | 2023-03-28 | 成都信息工程大学 | Cervical single cell image data generation method based on generation countermeasure network |
| TR202019150A1 (en) * | 2020-11-27 | 2022-06-21 | Ondokuz Mayis Ueniversitesi | Cerebral palsy detection system in newborn infants. |
| CN113197570B (en) * | 2021-05-07 | 2022-10-25 | 重庆大学 | A knee-crawling motion posture analysis system for infants and young children to assist in the diagnosis of cerebral palsy |
| CN114358194B (en) * | 2022-01-07 | 2024-11-19 | 吉林大学 | Abnormal limb behavior detection method for autism spectrum disorder based on posture tracking |
| CN114999648B (en) * | 2022-05-27 | 2023-03-24 | 浙江大学医学院附属儿童医院 | Early screening system, equipment and storage medium for cerebral palsy based on baby dynamic posture estimation |
Family Cites Families (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7539532B2 (en) * | 2006-05-12 | 2009-05-26 | Bao Tran | Cuffless blood pressure monitoring appliance |
| US9232912B2 (en) * | 2010-08-26 | 2016-01-12 | The Regents Of The University Of California | System for evaluating infant movement using gesture recognition |
| US11074495B2 (en) * | 2013-02-28 | 2021-07-27 | Z Advanced Computing, Inc. (Zac) | System and method for extremely efficient image and pattern recognition and artificial intelligence platform |
| US9700222B2 (en) * | 2011-12-02 | 2017-07-11 | Lumiradx Uk Ltd | Health-monitor patch |
| CN104083173B (en) * | 2014-07-04 | 2015-11-18 | 吉林大学 | A kind of quadruped movement observations system |
| US20160095524A1 (en) * | 2014-10-04 | 2016-04-07 | Government Of The United States, As Represented By The Secretary Of The Air Force | Non-Contact Assessment of Cardiovascular Function using a Multi-Camera Array |
| CN104765959A (en) * | 2015-03-30 | 2015-07-08 | 燕山大学 | Computer vision based evaluation method for general movement of baby |
| WO2016164373A1 (en) * | 2015-04-05 | 2016-10-13 | Smilables Inc. | Wearable infant monitoring device and system for determining the orientation and motions of an infant |
| CN105426919B (en) * | 2015-11-23 | 2017-11-14 | 河海大学 | The image classification method of non-supervisory feature learning is instructed based on conspicuousness |
| CN106230959A (en) * | 2016-08-09 | 2016-12-14 | 西安科技大学 | A kind of intelligence baby monitoring system |
| CN107679462B (en) * | 2017-09-13 | 2021-10-19 | 哈尔滨工业大学深圳研究生院 | A wavelet-based deep multi-feature fusion classification method |
| CN107684430A (en) * | 2017-09-29 | 2018-02-13 | 上海市上海中学 | Correcting device and its application method are detected based on Curie modules human body attitude |
| CN108764281A (en) * | 2018-04-18 | 2018-11-06 | 华南理工大学 | A kind of image classification method learning across task depth network based on semi-supervised step certainly |
| CN108836342A (en) * | 2018-04-19 | 2018-11-20 | 北京理工大学 | It is a kind of based on inertial sensor without feature human motion identification method |
| CN108567431A (en) * | 2018-05-14 | 2018-09-25 | 浙江大学 | A kind of intelligent sensing boots for measuring body gait and leg speed |
| CN108805188B (en) * | 2018-05-29 | 2020-08-21 | 徐州工程学院 | An Image Classification Method Based on Feature Recalibration Generative Adversarial Networks |
| CN108703760A (en) * | 2018-06-15 | 2018-10-26 | 安徽中科智链信息科技有限公司 | Human motion gesture recognition system and method based on nine axle sensors |
-
2018
- 2018-12-07 CN CN201811494749.0A patent/CN109620244B/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN109620244A (en) | 2019-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109620244B (en) | A Conditional Generative Adversarial Network and SVM-Based Approach for Infant Abnormal Behavior Detection | |
| Pang et al. | Automatic detection and quantification of hand movements toward development of an objective assessment of tremor and bradykinesia in Parkinson's disease | |
| CN110414306B (en) | Baby abnormal behavior detection method based on meanshift algorithm and SVM | |
| US20190223748A1 (en) | Methods and apparatus for mitigating neuromuscular signal artifacts | |
| CN106384093B (en) | A kind of human motion recognition method based on noise reduction autocoder and particle filter | |
| US20200160044A1 (en) | Physical activity quantification and monitoring | |
| CN118213039B (en) | Rehabilitation training data processing method and system based on deep reinforcement learning | |
| CA2907411A1 (en) | Feature extraction and classification to determine one or more activities from sensed motion signals | |
| CN110084286A (en) | A kind of human motion recognition method of sensor-based ECOC technology | |
| CN112438741B (en) | A driving state detection method and system based on EEG feature transfer learning | |
| CN112541870A (en) | Video processing method and device, readable storage medium and electronic equipment | |
| US20230367398A1 (en) | Leveraging machine learning and fractal analysis for classifying motion | |
| CN114027786A (en) | Sleep-disordered breathing detection method and system based on self-supervised memory network | |
| Chen et al. | An interpretable deep learning optimized wearable daily detection system for Parkinson’s disease | |
| Meena et al. | An explainable self-attention-based spatial–temporal analysis for human activity recognition | |
| Wang et al. | WCFormer: An interpretable deep learning framework for heart sound signal analysis and automated diagnosis of cardiovascular diseases | |
| CN116740426A (en) | A classification and prediction system for functional magnetic resonance images | |
| CN118230903B (en) | Accurate nutrition feeding method and system for pig feed based on image recognition | |
| CN113197558A (en) | Heart rate and respiratory rate detection method and system and computer storage medium | |
| CN114999648B (en) | Early screening system, equipment and storage medium for cerebral palsy based on baby dynamic posture estimation | |
| CN116869516A (en) | A comprehensive motion assessment method and system based on multi-source heterogeneous data | |
| CN113537489B (en) | Elbow angle prediction method, terminal equipment and storage medium | |
| US20250275700A1 (en) | Method and system for monitoring cognitive state of an individual | |
| EP4086811A1 (en) | Problematic behavior classification system and method based on deep neural network algorithm | |
| CN111062352B (en) | Method and server for identifying gait of terminal user based on mobile terminal data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210730 Termination date: 20211207 |