CN108416276B - Abnormal gait detection method based on human lateral gait video - Google Patents

Abstract

The invention discloses an abnormal gait detection method based on a human lateral gait video, which is characterized in that the normal gait video is collected by a common camera or a mobile phone, the stride change characteristic and the forward leaning characteristic of a human body are extracted from the side gait video, and a single classification support vector machine model is trained by using the characteristic information of the normal gait, so that the corresponding gait can be rapidly and effectively detected to be normal or abnormal. The invention does not need special detection equipment and gait detection in a specific place, and has high flexibility. The invention considers the forward leaning characteristic of the body when extracting the gait characteristics through the gait video, can effectively avoid the detection capability of the abnormal gait detection model from being influenced by abnormal training data, and improves the detection accuracy. The invention only needs to adopt the characteristic training single-classification support vector machine model of normal gait information, has few training samples, and can quickly and accurately detect normal/abnormal gait, thereby assisting a doctor to diagnose abnormal gait and improving the working efficiency of the doctor.

Description

An abnormal gait detection method based on human side gait videos

technical field

The invention relates to the field of gait recognition, in particular to a method for extracting features from a human side gait video and combining an abnormal gait detection model to realize abnormal gait detection.

Background technique

Gait is the walking posture of a person, and abnormal gait is the abnormal walking posture that occurs when a person's body is abnormal. Common causes of abnormal gait include pain, central nervous system abnormalities, and musculoskeletal injuries. There are many types of abnormal gaits. Typical abnormal gaits include spastic hemiplegic gait, spastic paraplegic gait, sensory ataxia gait, panic gait, myopathic gait, cross-threshold gait, and hysterical gait. Wait. The appearance of some typical gaits reflects the existence of characteristic diseases. By observing and analyzing abnormal gaits, symptoms can be diagnosed in patients, such as panic gait and freezing gait common in patients with Parkinson's disease. Therefore, the detection of abnormal gait is an important basis for doctors to diagnose. Although there are existing studies on human gait, they all focus on the use of gait for identification, and there are few studies on abnormal gait detection, and these methods are all based on normal human gait for identification and classification. , not suitable for general abnormal gait detection.

In clinical medicine, the detection of abnormal gait in patients often adopts direct observation or quantitative analysis based on objectively collected data such as electromyography. The collection and analysis of abnormal gait are usually carried out using external hardware instruments, such as the use of angle measurement devices to comprehensively collect the motion trajectories of major joints to analyze the regularity of human gait cycles; or to collect plantar pressure through a wearable gait collector. The abnormal gait in the categories of tiptoe walking, heel walking, outer stance and inner stance were analyzed according to the changes of plantar pressure data. These methods analyze the patient's gait based on hardware equipment, which requires high equipment, needs to be performed in a specific place, and has low flexibility, which is not conducive to the detection of abnormal gait in daily life. Individual studies (such as the invention patent "Abnormal Gait Detection Method and Abnormal Gait Detection System" with the application number of 2017107435559) classify normal abnormalities by extracting features from human contours, but only consider the characteristics of changes in the lower body stride, and do not consider some The abnormal gait is characterized by the forward leaning of the upper body in the panic gait. At the same time, the training model used is easily affected by the abnormal gait data in the training data, and cannot effectively detect the rich and diverse abnormal gait.

Therefore, there is a need for improvements to the prior art.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to propose an abnormal gait detection method based on human side gait video without professional detection equipment.

In order to solve the above problems, the present invention proposes a method for detecting abnormal gait based on human side gait video, comprising the following steps:

Step 1. Obtain the side gait video, and extract the silhouette contour image sequence from the side gait video;

Step 2. According to the silhouette image sequence obtained in step 1, construct the parameter sequence A of step width and height ratio, the x-axis horizontal distance between the center of the head and the feet and the height ratio forward leaning parameter sequence B, and by the parameter sequence A and parameter sequence B to extract features describing gait information;

Step 3, collecting the side gait video of normal gait as a training sample; the training sample sequentially extracts the features of its gait information through steps 1 and 2, and the extracted features are input into a single-classification support vector machine for training, Build an abnormal gait detection model;

Step 4: Obtain the side gait video to be detected, extract features of its gait information through steps 1 and 2 in turn, and input the features into the abnormal gait detection model constructed in step 3 for detection.

As an improvement on the abnormal gait detection method based on human side gait videos:

The step 2 includes the following steps:

2.1. Divide each frame of the silhouette silhouette image in the silhouette silhouette image sequence obtained in step 1 according to the pixel value of each point to obtain a human body area image;

2.2. Extract the height h and the x-axis coordinate of the head vertex x ₁ according to the human body region image obtained by segmentation in step 2.1;

2.3. Intercept the image of the foot area from 4/5h to the height of h of the human body area image obtained in step 2.1 according to the direction from top to bottom, and obtain the actual step width w and the x-axis coordinate x ₂ of the center point of both feet;

2.4. Calculate the parameters α and β according to the parameters calculated in steps 2.3 and 2.4:

Calculate the step width to height ratio parameter α, the calculation formula is

Calculate the horizontal forward tilt distance of the head and the height ratio parameter β, and the calculation formula is:

2.5. Add the parameters α and β extracted from each frame of silhouette image to parameter sequence A and parameter sequence B, respectively, until all silhouette images in the silhouette image sequence are processed, and obtain parameter sequence A and parameter sequence B; that is, , the parameter α is added to the parameter sequence A in sequence, and the parameter β is sequentially added to the parameter sequence B.

2.6. Extract the gait cycle, first half cycle, second half cycle, peak amplitude, trough amplitude, peak variance, trough-to-peak transition time and peak-to-trough gait from the curve drawn by the parameter sequence A obtained in step 2.5. 8 characteristics of transition time;

2.7. From the parameter sequence B obtained in step 2.5, extract four features representing the mean, variance, maximum and minimum values of forward leaning features;

2.8. Combine the features extracted in steps 2.6 and 2.7 to form the final feature G.

As a further improvement of the abnormal gait detection method based on human side gait videos:

The step 3 includes the following steps:

3.1. Collect side gait videos of normal gait to construct training samples;

3.2. The training samples constructed in step 3.1 are sequentially extracted from the features of their gait information through steps 1 and 2 to obtain a feature parameter training set;

3.3. Use the feature parameter training set obtained in step 3.2 to train the abnormality detection model, and construct the abnormal gait detection model.

As a further improvement of the abnormal gait detection method based on human side gait videos:

The abnormal gait detection model is a single-classification support vector machine.

As a further improvement of the abnormal gait detection method based on human side gait videos:

The optimization goal of the single-class SVM model is to find a minimum spherical surface with a center o and a radius R. The formula is as follows:

and meet the conditions:

Among them, u is the input data, i=1,...,N, N is the number of input data, C is the penalty factor, and ξ is the relaxation factor.

Using the Lagrange multiplier method and simplifying the constraints, the optimization problem can be transformed into:

The constraints are:

Among them, i=1,...,N, j=1,...,N, N is the number of input data, λ is the Lagrange multiplier.

The kernel function of this model adopts the RBF radial basis function:

K(u _i , u _j )=exp(-γ|u _i -u _j | ² ),

Among them, γ is the kernel parameter, and the value of this model is 0.083.

Compared with the prior art, the technical advantages of the present invention are:

1. The present invention extracts the step width-to-height ratio parameter (ie, parameter α) and the body forward leaning parameter (ie, parameter β) from the pedestrian profile as gait features, which effectively describes the gait features, wherein the step width is extracted and intercepted. The human foot area is reduced, to a certain extent, the interference of clothing such as coats on the step width extraction is reduced; by extracting the body forward leaning feature, it can better describe whether there is a forward leaning phenomenon during walking, which is helpful for abnormal Gait detection.

2. According to the selected gait characteristics, the present invention uses the normal gait data to perform model training with a single-classification support vector machine, and establishes an abnormal gait detection model, which can effectively distinguish the normal gait without using various abnormal gaits as training data. Gait and various abnormal gaits, to detect whether the gait to be tested is abnormal or normal.

Description of drawings

The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

Fig. 1 is the model building block diagram of the abnormal gait detection method based on the human side gait video of the present invention;

Fig. 2 is the schematic flow chart of extracting feature steps in Fig. 1;

3 is a schematic diagram of extracting parameters h and x ₁ from the human body contour in Embodiment 1;

4 is a schematic diagram of extracting parameters w and x ₂ from the human body contour in Embodiment 1;

Fig. 5 is the characteristic parameter α change curve diagram of normal gait and abnormal gait;

FIG. 6 is a graph showing the change of characteristic parameter β of normal gait and abnormal gait.

Detailed ways

The present invention will be further described below with reference to specific embodiments, but the protection scope of the present invention is not limited thereto.

Embodiment 1. An abnormal gait detection method based on human side gait video, as shown in Figure 1-6, by collecting side gait videos and effectively extracting gait parameters to describe the person's pace, stride, whether Forward leaning and other features to detect abnormal gait. As shown in Figure 1, it specifically includes the following steps:

Step 1. Use a common camera or mobile phone to shoot the side gait video of the person. When shooting the video, you need to use a tripod to fix the shooting device. The video format can be MP4 or AVI, and the side of the person is shot, including at least two gait cycles. A complete gait cycle refers to the process from a person's heel strike on one side to the same side's heel strike again. Then, based on the background difference method (the existing foreground extraction algorithm), the silhouette contour image sequence is extracted from the side gait video.

Step 2. As shown in Figure 2, according to the silhouette image sequence, build a parameter sequence A of the step width and height ratio, and a parameter sequence B of the horizontal distance of the x-axis between the center of the head and the two feet and the height ratio forward leaning. Sequence A and parameter sequence B extract 12 features to describe gait information. This feature extraction method has the advantage of effectively reflecting whether the pace time and distance parameters of a person are normal in the process of walking, and whether the body is leaning forward. Follow the steps below to process each frame of silhouette image sequence in sequence:

2.1. Segmenting the human body region: segment the human body region according to the pixel value of each point in each frame of the silhouette contour image in the sequence of silhouette contour images obtained in step 1.

The human body area segmentation adopts the image traversal method, and defines four variables, x_left, x_right, y_top and y_bottom, which respectively represent the x-axis coordinates of the leftmost point, the x-axis coordinate of the rightmost point, the y-axis coordinate of the highest point and the lowest point of the human body area on the original image. y-axis coordinate. Calculate the sum of the pixel values of each image column or row in turn according to the four directions from left to right, right to left, top to bottom, and bottom to top. Since in the binarized image, the value of the black area is 0, and the value of the white area is 255. Therefore, when the sum of the calculated column or row pixel values is not 0, it means that the boundary line of the human body area is detected, the variable value is recorded, and the traversal is terminated. Find.

The specific method for calculating x_left in this embodiment is: initialize x_left=0, calculate the sum of pixel values of each image column in turn from left to right, when the sum of pixel values is 0, add 1 to the x_left value, and continue to calculate the following The sum of the pixel values of an image column is stopped until the sum of the pixel values is not 0. At this time, the pixel point whose pixel value is not 0 is the leftmost point, and its x-axis coordinate is the x_left at this time, and x_left is recorded; calculation The method of x_right is: initialize x_right as the size and width of the image (such as 300), and then calculate the sum of each pixel column column by column in the direction from right to left. When the sum of the pixel values is 0, the x_right value is reduced by 1, and continue Calculate the sum of the pixel values of the next image column until the sum of the pixel values is not 0. At this time, the pixel point whose pixel value is not 0 is the rightmost point, and its x-axis coordinate is the x_right at this time. Record x_right ; The method of calculating y_top and y_bottom is the same as the above calculation method, so the description will not be repeated.

Finally, according to the four calculated variables (x_left, x_right, y_top and y_bottom), the silhouette contour image is cropped to obtain the human body area image, as shown in Figure 3.

2.2, according to the human body region image obtained by step 2.1, extract the height h, the head vertex x-axis coordinate x ₁ ;

The height h is the size and height of the image of the human body area, that is, the difference between y_bottom and y_top in step 2.1.

The calculation method of x ₁ is: initialize x ₁ =0, traverse each pixel in the first line of the human body area image, if the pixel value is 0, add 1 to the x value, continue to check the next pixel, until the pixel is not Terminate at 0, record the x ₁ value at this time.

2.3. Take the image of the human body area obtained in step 2.1 according to the direction from top to bottom to intercept the image of the foot area from 4/5h to the height of h (that is, intercept the height from 0 to 1/5h in the direction from bottom to top) image of the foot area), thereby reducing the interference of clothing such as coats on the step width extraction. As shown in FIG. 4 , the actual stride width w and the x-axis coordinate x ₂ of the center point of the two feet are obtained from the intercepted image of the foot area.

The calculation method of w and x ₂ is: define two variables x ₁ _left and x ₁ _right, which respectively represent the x-axis coordinate of the leftmost point and the x-axis coordinate of the rightmost point in the foot area. Calculate the sum of the pixel values of each image column in turn according to the left-to-right and right-to-left directions.

The method for calculating x ₁ _left in this embodiment is: initialize x ₁ _left, calculate the sum of pixel values of each image column in turn from left to right, when the sum of pixel values is 0, add 1 to the value of x ₁ _left, Continue to calculate the sum of the pixel values of the next image column until the sum of the pixel values is not 0. At this time, the pixel point whose pixel value is not 0 is the leftmost point of the foot area, and its x-axis coordinate is recorded in x ₁ _left; the calculation of x ₁ _right is obtained in the direction from right to left by this method (same as the calculation method of x_right in step 2.1).

计算。w is calculated according to the calculation formula w=|x ₁ _right-x ₁ _left|, and x ₂ is calculated according to the formula

calculate.

2.4. Calculate the parameters α and β according to the parameters calculated in steps 2.3 and 2.4 (ie h, x ₁ , w and x ₂ ):

计算步宽与身高比参数α，自定义公式

计算头部水平前倾距离与身高比参数β(即身体前倾参数)；by custom formula

Calculate the step width to height ratio parameter α, custom formula

Calculate the horizontal forward tilt distance of the head to the height ratio parameter β (that is, the body forward tilt parameter);

2.5. Add the parameter α extracted from each frame of silhouette image to the parameter sequence A, and the parameter β to the parameter sequence B, until all silhouette images in the silhouette image sequence are processed, and obtain the final parameter sequence A and B .

In this embodiment, a camera is used to record videos of a normal gait and an abnormal gait (panic gait) respectively, and processing is performed according to the above steps to obtain corresponding parameter sequences A and B. As shown in Figure 5, the left picture is the graph corresponding to the parameter sequence A of the normal gait, and the right graph is the graph corresponding to the abnormal gait parameter sequence A, the abscissa is the frame number, and the ordinate is the parameter α; 5 It can be seen that the gait movement of normal people is periodic. As the walking forward and the left and right feet switch, the gait parameter α fluctuates between 0.15 and 0.50, and the peak and trough amplitudes are basically stable; the panic gait parameter α The value fluctuates between 0.15 and 0.25, and the peak amplitude is relatively low, reflecting the characteristics of patients with small steps and difficulty in walking.

As shown in Figure 6, the left picture is the graph corresponding to the parameter sequence B of normal gait, the right graph is the graph corresponding to the abnormal gait parameter sequence B, the abscissa is the frame number, and the ordinate is the parameter β; It can be seen that the parameter β value of the normal gait is roughly between 0.0 and 0.1, and the fluctuation is not large; the parameter β value of the panic gait is between 0.1 and 0.4, the fluctuation is large, and the average value is higher than that of the normal gait. It can be seen from Figure 5 and Figure 6 that the parameter sequences A and B can effectively distinguish the normal gait from the abnormal gait.

2.6. Extract the gait cycle G1, the first half cycle G2, the second half cycle G3, the peak amplitude G4, the trough amplitude G5, the peak variance G6, and the transition from the trough to the peak from the curve drawn by the parameter sequence A obtained in step 2.5. The time G7 and the peak-to-valley transition time G8 have a total of 8 features.

As shown in FIG. 5 , the gait cycle G1 is defined as the number of frames between the i-th peak and the i+2-th peak, which is equivalent to the time of one gait cycle. The larger the number of frames, the longer the cycle, the slower the walking speed, and the possibility of abnormal gait. The first half cycle of gait G2 is defined as the number of frames between the i-th peak to the i+1-th peak. This feature should be roughly equal to the G3 feature under normal walking conditions. When the difference between the two is large, there may be a single leg abnormality. The first half cycle of gait G3 is defined as the number of frames between the i+1th peak to the i+2th peak. The peak amplitude G4 is the highest point of the curve, which actually reflects the step length of a person. When the value is small, it means that the pace is small, and there may be difficulty in stepping. The trough amplitude G5 is the lowest point of the curve, which actually reflects the moment when a person's legs overlap during walking. When the value is large, there may be difficulty in walking with the aid of a support. The peak variance G6 is the variance of all the peaks in the gait video time, which actually reflects the change of the stride during walking. The larger the variance, the more abnormal the patient's walking and the uneven stride length. The transition time G7 from the trough to the peak is the time from the overlapping of the legs to the landing of both feet, and the transition time G8 from the trough to the peak is the time from the landing of both feet to the overlapping of the legs. pace.

2.7. Extract the body forward leaning feature from the parameter sequence B obtained in step 2.5, namely the mean value G9, variance G10, maximum value G11 and minimum value G12 of the parameter β, a total of 4 features, which can be used to describe the walking process of people through these features. Whether there is a body leaning forward or rocking back and forth.

As shown in FIG. 6 , the maximum value G11 is the highest point of the curve, and the minimum value G12 is the highest point of the curve.

Note: The calculation of the above-mentioned mean value G9 and variance G10 is the prior art, so the calculation method is not described in detail in this specification.

2.8. The features extracted in steps 2.6 and 2.7 are combined to form the final feature G as follows:

G=[G1 G2 G3 G4 G5 G6 G7 G8 G9 G10 G11 G12 ]

Step 3, collect the side gait videos of normal gait (hereinafter referred to as normal gait videos) as training samples; the training samples sequentially extract the features of its gait information through steps 1 and 2, and input the features into a single The classification support vector machine is trained to construct an abnormal gait detection model. This model building method has the advantage of effectively distinguishing between normal gaits and various abnormal gaits without using various abnormal gaits as training data. This step specifically includes the following steps:

3.1. Collect normal gait videos and side gait videos of abnormal gait (hereinafter referred to as abnormal gait videos); randomly select 70% of the collected normal gait videos to construct training samples, and use the remaining 30% of normal gait videos to construct training samples. The gait videos were mixed with the collected abnormal gait videos to form the test samples.

3.2. The training samples constructed in step 3.1 are sequentially extracted from the features of the lateral gait video (ie, the features of normal gait information) through steps 1 and 2 to obtain a feature parameter training set.

3.3. The feature parameter training set is input into the single-classification support vector machine for training, and the abnormal gait detection model is obtained by training.

One-class support vector machine (One-class-SVM) is an existing anomaly detection model, which is suitable for single-classification tasks such as positive and negative sample cloth imbalance and anomaly detection. In addition to single-classification support vector machines, common anomaly detection models also include models such as isolation forest and local outlier factor detection. The main idea of single-class SVM is to train a hypersphere model including all positive examples. When a new data needs to be judged, if the data point falls within the hypersphere, it belongs to this class, and if it falls outside the sphere , it does not belong, it is abnormal data. The present invention adopts the existing single-classification support vector machine, uses the feature parameter training set to train the single-classification support vector machine, and constructs a hypersphere, the hypersphere surrounds the features containing normal gait information, so as to realize the abnormal gait. detection. The optimization goal of the model is to find a smallest spherical surface with center o and radius R:

and meet the conditions:

Among them, u is the input data, i=1,...,N, N is the number of input data, C is the penalty factor, ξ is the relaxation factor, and F is the objective function.

The Lagrange multiplier method is used to solve the problem and simplify the constraints (Note: The Lagrange multiplier method is a commonly used method for solving optimization problems and is an existing mathematical method, so its calculation method is not described in detail in this manual) , the optimization problem can be transformed into:

The constraints are:

Among them, i=1,...,N, j=1,...,N, N is the number of input data, λ is the Lagrangian multiplier, and L is the Lagrangian function.

The kernel function of this model adopts the RBF radial basis function:

K(u _i , u _j )=exp(-γ|u _i -u _j | ² ),

Among them, γ is the kernel parameter, the value of this model is 0.083, and K is the radial basis function.

3.4. Extract the features of its gait information from the test sample (or the actual side gait video to be detected) constructed in step 3.1 through steps 1 and 2 in turn, and input the features into the abnormal gait detection model constructed in step 3.3, If the data feature points fall within the training hypersphere, the output +1 is the normal gait, otherwise the output -1 is the abnormal gait. Thereby, it can quickly detect whether the gait is abnormal.

Experiments, the effect of forward leaning features on the accuracy of abnormal gait detection:

The test sample in this experiment consists of 85 normal gait videos and 35 abnormal gait videos, of which 35 abnormal gait videos consist of 10 lameness, 10 spastic hemiplegia, and 5 sensory ataxia. gait and 10 panicked gaits. After sequentially extracting the final feature G of each side gait video in the test sample, input the feature G into the abnormal gait detection model established in Example 1 for detection, and the detection results are shown in Table 1:

Table 1

Comparative example, cancel step 2.7 in the embodiment, and the rest are identical to the embodiment; that is, the final feature G only includes the features G1-G8 (cancel the mean value G9, variance G10, maximum value G11 and minimum value G12); adopt the same in the embodiment The training samples are extracted, and the final feature G is extracted to train a single-class SVM to construct an abnormal gait detection model; then the final feature G of each side gait video in the test sample is extracted in turn, and the feature G is input to the comparison scale The established abnormal gait detection model is used for detection, and the detection results are shown in Table 1.

It can be seen from Table 1 that the accuracy of the model detection results is greatly improved after adding body forward leaning features (mean value G9, variance G10, maximum value G11 and minimum value G12). Judging from the detection results of normal gait data, the accuracy of correct detection of positive samples has been improved compared with that before the body forward leaning feature was added, which reduced misjudgment; from the detection results of abnormal gait data, the increase of body forward leaning After the feature, the accuracy rate of correct detection as a negative sample reaches 100.0%, in particular, the model detection effect has been significantly improved for the panic gait with the body forward leaning feature.

To sum up, the present invention collects side gait videos through common ordinary cameras or mobile phones, extracts human stride variation features and forward leaning features from them, and uses the feature information of normal gait to train a single-classification support vector machine model, which can be Quickly and effectively detect whether the corresponding gait is normal or abnormal. The invention does not need special detection equipment, does not need to perform gait detection in a specific place, and has high flexibility. The present invention considers the body forward leaning feature when extracting the gait feature from the gait video, and can effectively avoid the detection ability of the abnormal gait detection model from being affected by the abnormal training data, thereby improving the detection accuracy. The present invention only needs to use the features of normal gait information to train a single-classification support vector machine model, with few training samples, and can quickly and accurately detect normal/abnormal gait, thereby assisting doctors in diagnosing abnormal gaits and improving doctors' work efficiency.

Finally, it should also be noted that the above enumeration is only a few specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and many modifications are possible. All deformations that those of ordinary skill in the art can directly derive or associate from the disclosure of the present invention shall be considered as the protection scope of the present invention.

Claims (3)

1. The abnormal gait detection method based on the human lateral gait video is characterized by comprising the following steps:

step 1, acquiring a lateral gait video, and extracting a silhouette image sequence from the lateral gait video;

shooting the side of a person, wherein the shooting comprises at least two gait cycles; a complete gait cycle refers to the process that a person lands from the heel at one side to the heel at the same side again; extracting a silhouette image sequence from the side gait video based on a background subtraction method;

step 2, according to the silhouette image sequence obtained in the step 1, constructing a parameter sequence A of step width to height ratio, a parameter sequence B of x-axis horizontal distance between the head and the centers of two feet and height ratio forward inclination, and extracting features describing gait information from the parameter sequence A and the parameter sequence B;

the method comprises the following steps:

2.1, dividing each frame of silhouette image in the silhouette image sequence obtained in the step 1 according to pixel values of each point to obtain a human body area image;

2.2, extracting the height h according to the human body region image obtained by segmentation in the step 2.1, and extracting the x-axis coordinate x of the vertex of the head₁；

2.3, intercepting the foot area image from 4/5h to h height of the human body area image obtained in the step 2.1 from top to bottom, and obtaining the actual step width w and the x-axis coordinate x of the center point of two feet₂；

2.4, calculating parameters alpha and beta according to the parameters calculated in the step 2.3 and the step 2.4:

calculating the step width-height ratio parameter alpha according to the formula

Calculating the parameter beta of the horizontal forward tilting distance of the head and the height ratio, wherein the calculation formula is

2.5, adding the parameters alpha and beta extracted from each frame of silhouette image into the parameter sequence A and the parameter sequence B respectively until all the silhouette images in the silhouette image sequence are processed, and obtaining the parameter sequence A and the parameter sequence B;

2.6, extracting 8 characteristics of the gait cycle, the first half cycle, the second half cycle, the peak amplitude, the trough amplitude, the peak variance, the transition time from the trough to the peak and the transition time from the peak to the trough from the curve drawn by the parameter sequence A obtained in the step 2.5;

2.7, extracting 4 features of mean value, variance, maximum value and minimum value representing the anteversion features of the body from the parameter sequence B obtained in the step 2.5;

2.8, integrating the features extracted in the steps 2.6 and 2.7 to form a final feature G;

step 3, collecting a side gait video of a normal gait as a training sample; the training sample sequentially extracts the features of the gait information through the step 1 and the step 2, and inputs the extracted features into a single classification support vector machine for training to construct an abnormal gait detection model;

the method comprises the following steps:

3.1, collecting a side gait video of a normal gait to construct a training sample;

3.2, extracting the features of the gait information of the training sample constructed in the step 3.1 sequentially through the step 1 and the step 2 to obtain a feature parameter training set;

3.3, training an abnormal gait detection model by adopting the characteristic parameter training set obtained in the step 3.2, and constructing the abnormal gait detection model;

the abnormal gait detection model is a single classification support vector machine;

and 4, acquiring a side gait video to be detected, sequentially extracting the features of the gait information through the step 1 and the step 2, and inputting the features into the abnormal gait detection model constructed in the step 3 for detection.

2. The method of detecting abnormal gait based on human lateral gait video according to claim 1, characterized in that:

the optimization goal of the single-classification support vector machine model is to solve a minimum spherical surface with the center of o and the radius of R, and the formula is as follows:

and satisfy the conditions：(u_i-o)^T(u_i-o)≤R²+ξ_i

Wherein u is input data, i is 1, …, N is the number of input data, C is a penalty factor, ξ is a relaxation factor;

by adopting a Lagrange multiplier method and simplifying constraint conditions, the optimization problem can be converted into the following steps:

the constraint conditions are as follows:

wherein i is 1, …, N, j is 1, …, N is the number of input data, λ is lagrange multiplier;

the kernel function of the model adopts RBF radial basis function:

K(u_i，u_j)＝exp(-γ|u_i-u_j|²)，

wherein gamma is a nuclear parameter, and the value of the model is 0.083.

3. The abnormal gait detection method based on human lateral gait video according to claim 1 or 2, characterized in that:

step 2.1: the human body area segmentation adopts an image traversal method, defines four variables of x _ left, x _ right, y _ top and y _ bottom, and respectively represents the leftmost point x-axis coordinate, the rightmost point x-axis coordinate, the highest point y-axis coordinate and the lowest point y-axis coordinate of the human body area on an original image; sequentially calculating the sum of pixel values of each image column or row according to four directions of left-to-right, right-to-left, top-to-bottom and bottom-to-top respectively; in the binary image, the black area value is 0, and the white area value is 255, so when the sum of the pixel values of the calculation columns or rows is not 0, the boundary line of the human body area is detected, the variable value is recorded, and traversal search is terminated;

the specific method for calculating x _ left comprises the following steps: initializing x _ left to be 0, sequentially calculating the sum of pixel values of each image column from left to right, adding 1 to the x _ left value when the sum of the pixel values is 0, continuously calculating the sum of the pixel values of the next image column until the sum of the pixel values is not 0, stopping, taking the pixel point of which the pixel value is not 0 as the leftmost point at the moment, taking the x-axis coordinate of the pixel point as the x _ left at the moment, and recording the x _ left;

the method for calculating x _ right comprises the following steps: initializing x _ right as the size width of an image, calculating the sum of each pixel column row by row according to the direction from right to left, when the sum of the pixel values is 0, subtracting 1 from the x _ right value, continuously calculating the sum of the pixel values of the next image column row until the sum of the pixel values is not 0, stopping, wherein the pixel point of the pixel value of the row not being 0 is the rightmost point, the x-axis coordinate of the pixel value of the row not being 0 is the x _ right at the moment, and recording the x _ right;

finally, according to the four calculated variables (x _ left, x _ right, y _ top and y _ bottom), cutting the silhouette image to obtain a human body area image;

step 2.2:

the height h is the size height of the image of the human body region, i.e. the difference between y _ bottom and y _ top in step 2.1;

x₁the calculation method is as follows: initialization x₁When the pixel point value is 0, each pixel point of the first line of the human body region image is traversed, if the pixel point value is 0, the x value is added with 1, the next pixel point is continuously checked until the pixel point value is not 0, the operation is terminated, and the x at the moment is recorded₁A value;

step 2.3:

w and x₂The calculation method comprises the following steps: definition of x₁_left，x₁Two variables _right, which respectively represent the x-axis coordinate of the leftmost point and the x-axis coordinate of the rightmost point of the foot area; sequentially calculating the sum of pixel values of each image column according to a left-to-right direction and a right-to-left direction respectively;

calculating x₁The method of _ left is as follows: initialization x₁A left, calculating the sum of pixel values of each image column in turn from left to right, and obtaining the current imageWhen the sum of the prime values is 0, x₁Adding 1 to the _leftvalue, continuously calculating the sum of pixel values of the next image row until the sum of the pixel values is not 0, stopping calculating the sum of the pixel values until the sum of the pixel values is not 0, wherein the pixel point of the row with the pixel value not 0 is the leftmost point of the foot area, and the x-axis coordinate of the pixel point is recorded in the x-axis coordinate₁In _left; x is the number of₁The calculation of _ right is obtained in the right to left direction according to the method;

w is calculated according to the formula w ═ x₁_right-x₁Let | calculate, x₂According to the formula

Calculating;

step 2.5:

and respectively recording videos of normal gait and abnormal gait by using a camera, and performing corresponding processing to obtain corresponding parameter sequences A and B.

Images