CN108961675A - Fall detection method based on convolutional neural networks - Google Patents

Abstract

The present invention relates to a fall detection method based on a convolutional neural network, comprising: training the convolutional neural network, the training of the convolutional neural network specifically includes: preprocessing each frame of image obtained, the preprocessing work includes Foreground extraction, normalization, and whitening operations are performed in sequence; the ResNet network is pre-trained on the ImageNet dataset to obtain a pre-trained model. The classification method based on convolutional neural network is applied to the fall detection method. At the same time, in order to improve the accuracy of the system and reduce the complexity of the operation, an improved foreground detection method is used to extract the people in the complex background, and then the The processed images are put into the convolutional neural network for model training.

Description

Fall detection method based on convolutional neural network

technical field

The invention relates to a fall detection method, in particular to a fall detection method based on a convolutional neural network.

Background technique

Nowadays, the aging trend of the society is getting worse, the decline of the physical function of the elderly and the more and more common phenomenon of living alone make falls one of the main reasons for the injury of the elderly, so it is of great significance to detect the fall behavior.

Traditional computer vision-based fall detection methods usually extract features manually, which requires a huge amount of engineering, poor generalization ability, and low accuracy. Unlike traditional feature extraction methods, convolutional neural networks can automatically extract features, and the trained model has geometric invariance, which can overcome problems caused by changes in lighting and shooting angles.

The traditional technology has the following technical problems:

The current fall detection system is mainly divided into two categories: the first is the sensor-based wearable detection system; the other is the video-based detection system. At present, research on wearable recognition systems based on three-dimensional acceleration or torso angular velocity has been relatively mature. However, wearable devices generally need to be worn on the neck or waist, and wearing them for a long time will make users feel uncomfortable. The vision-based detection system captures the movement of the target through one or several cameras, and uses a specific image processing algorithm to determine the image features when a fall occurs, thereby distinguishing the fall from daily activities. At present, the commonly used vision-based fall detection algorithms are mainly threshold method and intelligent algorithm. The threshold method usually detects the head position or the center of gravity of the human body. Diraco considers a fall when it judges that the center of the human body is lower than the specified height and maintains it for more than 4s. Rougier et al. determine whether a fall has occurred by locating the head position, estimating the head position in the next frame image through a particle filter, calculating the horizontal and vertical speeds, and comparing with the threshold. These methods are simple to implement, but the accuracy is easily affected by external factors such as the environment. The method based on machine learning is mainly to extract people from the image first, then manually extract features, and input the obtained features into the model to realize the detection and recognition of falling behavior. This method requires manual extraction of features, the amount of work is huge, and most of them only stay on the binary classification problem. Considering that the requirements for smart homes will become higher in the future, the recognition of various postures of the human body has become an indispensable part. .

Contents of the invention

Based on this, it is necessary to provide a fall detection method based on convolutional neural network for the above technical problems, and apply the classification method based on convolutional neural network to the fall detection method. At the same time, in order to improve the accuracy of the system and reduce the calculation The complexity, using an improved foreground detection method to extract the characters in the complex background, and then put the processed image into the convolutional neural network for model training.

A fall detection method based on convolutional neural network, comprising:

Training convolutional neural network, the training convolutional neural network specifically includes:

Perform preprocessing on each frame of image obtained, and the preprocessing work includes foreground extraction, normalization, and whitening operations in sequence;

Pre-train the ResNet network on the ImageNet dataset to obtain a pre-trained model;

The step "preprocess each frame of image obtained, the preprocessing work includes foreground extraction, normalization, and whitening operations in turn;" the processed picture is put into the pre-training model for model training, and the obtained the parameters of the model; and

Input the test set into the trained model, and use the test set to test the accuracy of the model;

Use the trained convolutional neural network to detect pictures;

Wherein, the method for foreground extraction specifically includes:

Image processing using background subtraction;

Image processing using a mixture of Gaussian models;

Sum the result of processing the image using the background subtraction method and the result of processing the image using the mixture Gaussian model.

The above fall detection method based on convolutional neural network applies the classification method based on convolutional neural network to the fall detection method. At the same time, in order to improve the accuracy of the system and reduce the complexity of the operation, an improved foreground detection method is used To extract the characters in the complex background, and then put the processed image into the convolutional neural network for model training.

In another embodiment, in the step "preprocessing each frame of image obtained, the preprocessing work includes foreground extraction, normalization, and whitening operations in sequence;" in each frame of image obtained It is obtained by reading the video file.

In another embodiment, after the step "using the trained convolutional neural network to detect pictures", the training convolutional neural network specifically further includes: displaying the detection effect diagram of each frame of pictures, and implementing the model The convolution kernel visualization.

In another embodiment, in the step "displaying the detection effect diagram of each frame of pictures, and realizing the visualization of the convolution kernel of the model", the detection effect diagram of each frame of pictures is displayed on the matlab platform, and the model is realized. Convolution kernel visualization.

In another embodiment, the step "processing the image using the background subtraction method;" specifically includes:

Take the average value of the previous few frame images and use it as the initial background image B _t ;

The grayscale of the previous frame image and the background image is subtracted, and its absolute value is N _t (x, y) The formula is

N _t (x,y)＝|I _t (x,y)-B _t (x,y)|

For the pixel (x, y) of the current frame, if there is |I _t (x, y)-B _t (x, y)|≥T, then the pixel point is the foreground point, that is, the current image frame is updated as;

Update the background image with the current frame image.

In another embodiment, the step "processing the image using a mixture of Gaussian models;" specifically includes:

When the Gaussian mixture model is used to model the background, the pixel value of each pixel of the sequence image can be simulated by k Gaussian models, so at time t, the probability density function of a certain pixel value can be expressed as:

Among them, w _{i, t} represents the weight of the Gaussian model, and the probability density function of the Gaussian model is expressed as:

Next, the K Gaussian mixture models are sorted according to the size of the quotient of the weight divided by the standard deviation, and then the previous B Gaussian models are selected for discrimination, where the value of B is expressed as:

Put each pixel in the new image frame into the sorted K Gaussian models for judgment, and the judgment conditions are:

||X _t -μ _t ||≤β∑ ^1/2

Among the previous B Gaussian models, if one of the Gaussian models meets the above conditions, then the pixel is judged as the background, and if none of the above conditions is satisfied in the B Gaussian models, the pixel is judged to belong to the foreground.

For each Gaussian model, assuming that the above conditions are not true, the weight of this Gaussian model must be reduced. If the above conditions are true, the Gaussian mixture model must be updated. The specific operation method is shown in the following formula:

w _i,t ＝(1-λ)w _i,t-1 +λBM _t

μ _i,t ＝(1-α)μ _i,t-1 +αX _i,t

∑ _i,t ＝(1-α)∑ _i,t-1 +α(X _i,t -μ _i,t )(X _i,t -μ _i,t ) ^T

α=λ/w _i,t

Among them, if the pixel is the foreground, BM=0, otherwise BM=1, and finally an initial Gaussian model is used to replace the Gaussian model with the smallest weight, the threshold T, the learning rate λ, and the parameter β are all constants specified in advance.

In another embodiment, in the step "input the test set into the trained model, and use the test set to test the accuracy of the model;", the test set comes from UR Fall Detection Dataset.

A computer device includes a memory, a processor, and a computer program stored in the memory and operable on the processor, and the processor implements the steps of any one of the methods when executing the program.

A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of any one of the methods described above are implemented.

A processor, the processor is used to run a program, wherein the program executes any one of the methods when running.

Description of drawings

FIG. 1 is a schematic flowchart of a fall detection method based on a convolutional neural network provided in an embodiment of the present application.

FIG. 2 is a schematic diagram of a residual learning building block in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 3 is a schematic diagram of a loss value function curve in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 4 is a flow chart of testing a model in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 5 is a schematic diagram of the effect of the background subtraction method in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 6 is a schematic diagram of the effect of a Gaussian mixture environment model in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 7 is a schematic diagram of the effect of an improved foreground detection method in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 8 is an RGB image of foreground extraction in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 9 is a visualization of convolution kernels in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 10 is a feature map of the first layer in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

FIG. 11 is a feature map of the second layer in a fall detection method based on a convolutional neural network provided by an embodiment of the present application.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Referring to Figure 1, a fall detection method based on convolutional neural network, including:

Training convolutional neural network, the training convolutional neural network specifically includes:

Perform preprocessing on each frame of image obtained, and the preprocessing work includes foreground extraction, normalization, and whitening operations in sequence;

Pre-train the ResNet network on the ImageNet dataset to obtain a pre-trained model;

The step "preprocess each frame of image obtained, the preprocessing work includes foreground extraction, normalization, and whitening operations in turn;" the processed picture is put into the pre-training model for model training, and the obtained the parameters of the model; and

Input the test set into the trained model, and use the test set to test the accuracy of the model;

Use the trained convolutional neural network to detect pictures;

Wherein, the method for foreground extraction specifically includes:

Image processing using background subtraction;

Image processing using a mixture of Gaussian models;

Sum the result of processing the image using the background subtraction method and the result of processing the image using the mixture Gaussian model.

The above fall detection method based on convolutional neural network applies the classification method based on convolutional neural network to the fall detection method. At the same time, in order to improve the accuracy of the system and reduce the complexity of the operation, an improved foreground detection method is used To extract the characters in the complex background, and then put the processed image into the convolutional neural network for model training.

In another embodiment, in the step "preprocessing each frame of image obtained, the preprocessing work includes foreground extraction, normalization, and whitening operations in sequence;" in each frame of image obtained It is obtained by reading the video file.

In another embodiment, after the step "using the trained convolutional neural network to detect pictures", the training convolutional neural network specifically further includes: displaying the detection effect diagram of each frame of pictures, and implementing the model The convolution kernel visualization.

In another embodiment, in the step "displaying the detection effect diagram of each frame of pictures, and realizing the visualization of the convolution kernel of the model", the detection effect diagram of each frame of pictures is displayed on the matlab platform, and the model is realized. Convolution kernel visualization.

In another embodiment, the step "processing the image using the background subtraction method;" specifically includes:

Take the average value of the previous few frame images and use it as the initial background image B _t ;

The grayscale of the previous frame image and the background image is subtracted, and its absolute value is N _t (x, y) The formula is

N _t (x,y)＝|I _t (x,y)-B _t (x,y)|

For the pixel (x, y) of the current frame, if there is |I _t (x, y)-B _t (x, y)|≥T, then the pixel point is the foreground point, that is, the current image frame is updated as;

Update the background image with the current frame image.

In another embodiment, the step "processing the image using a mixture of Gaussian models;" specifically includes:

When the Gaussian mixture model is used to model the background, the pixel value of each pixel of the sequence image can be simulated by k Gaussian models, so at time t, the probability density function of a certain pixel value can be expressed as:

Among them, w _{i, t} represents the weight of the Gaussian model, and the probability density function of the Gaussian model is expressed as:

Next, the K Gaussian mixture models are sorted according to the size of the quotient of the weight divided by the standard deviation, and then the previous B Gaussian models are selected for discrimination, where the value of B is expressed as:

Put each pixel in the new image frame into the sorted K Gaussian models for judgment, and the judgment conditions are:

||X _t -μ _t ||≤β∑ ^1/2

Among the previous B Gaussian models, if one of the Gaussian models meets the above conditions, then the pixel is judged as the background, and if none of the above conditions is satisfied in the B Gaussian models, the pixel is judged to belong to the foreground.

For each Gaussian model, assuming that the above conditions are not true, the weight of this Gaussian model must be reduced. If the above conditions are true, the Gaussian mixture model must be updated. The specific operation method is shown in the following formula:

w _i,t ＝(1-λ)w _i,t-1 +λBM _t

μ _i,t ＝(1-α)μ _i,t-1 +αX _i,t

∑ _i,t ＝(1-α)∑ _i,t-1 +α(X _i,t -μ _i,t )(X _i,t -μ _i,t ) ^T

α=λ/w _i,t

Among them, if the pixel is the foreground, BM=0, otherwise BM=1, and finally an initial Gaussian model is used to replace the Gaussian model with the smallest weight, the threshold T, the learning rate λ, and the parameter β are all constants specified in advance.

In another embodiment, in the step "input the test set into the trained model, and use the test set to detect the accuracy of the model;", the test set is from URFall Detection Dataset.

A computer device includes a memory, a processor, and a computer program stored in the memory and operable on the processor, and the processor implements the steps of any one of the methods when executing the program.

A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of any one of the methods described above are implemented.

A processor, the processor is used to run a program, wherein the program executes any one of the methods when running.

A specific application scenario of the present invention is given below:

1. Image preprocessing

In a specific embodiment, an improved foreground detection method is used to extract the foreground of a person. At present, the methods of foreground extraction mainly include frame difference method, background difference method, optical flow method, and mixed Gaussian model.

1) Background difference method

Wherein, if it is assumed that I _t and B _t are the current image frame and the background frame respectively, and T is the gray threshold value of the foreground detection, the algorithm steps of the background difference method are as follows:

1) Take the average value of the previous few frame images and use it as the initial background image B _t ;

2) Subtract the grayscale of the previous frame image and the background image, and take its absolute value as N _t (x, y); the formula is

N _t (x,y)＝|I _t (x,y)-B _t (x,y)|

3) For the pixel (x, y) of the current frame, if there is |I _t (x, y)-B _t (x, y)|≥T, then the pixel point is the foreground point, that is, update the current image frame as

4) Update the background image with the current frame image.

2) Mixed Gaussian model

The mixed Gaussian model is an adaptive mixed Gaussian background extraction method based on background modeling proposed by Stauffer et al. When the Gaussian mixture model is used to model the background, the pixel value of each pixel of the sequence image can be simulated by k Gaussian models, so at time t, the probability density function of a certain pixel value can be expressed as:

Among them, w _{i, t} represents the weight of the Gaussian model, and the probability density function of the Gaussian model can be expressed as:

Next, the K Gaussian mixture models are sorted according to the size of the quotient of the weight divided by the standard deviation, and then the previous B Gaussian models are selected for discrimination, where the value of B is expressed as:

Put each pixel in the new image frame into the sorted K Gaussian models for judgment, and the judgment conditions are:

||X _t -μ _t ||≤β∑ ^1/2

Among the previous B Gaussian models, if one of the Gaussian models meets the above conditions, then the pixel is judged as the background, and if none of the above conditions is satisfied in the B Gaussian models, the pixel is judged to belong to the foreground.

For each Gaussian model, assuming that the above conditions are not true, the weight of this Gaussian model must be reduced. If the above conditions are true, the Gaussian mixture model must be updated. The specific operation method is shown in the following formula:

w _i,t ＝(1-λ)w _i,t-1 +λBM _t

μ _i,t ＝(1-α)μ _i,t-1 +αX _i,t

∑ _i,t ＝(1-α)∑ _i,t-1 +α(X _i,t -μ _i,t )(X _i,t -μ _i,t ) ^T

α=λ/w _i,t

Wherein, if the pixel is foreground, BM=0, otherwise BM=1. Finally, an initial Gaussian model is used to replace the Gaussian model with the smallest weight. The threshold T, learning rate λ, and parameter β are all constants specified in advance.

3) Improved foreground detection method

Although the method of background subtraction is simple and the amount of calculation is small, it will produce "ghosting" phenomenon. The mixed Gaussian model not only models the background, but also models the foreground, so it is very sensitive to sudden changes in global brightness. In order to solve the limitations caused by using only one method, the present invention proposes an improved foreground detection method, that is, performing an "AND" operation on the processing effects of the two can well solve the problems of "ghosting" and light sensitivity . For the pixel point (x, y) of the image matrix, it is assumed that the output of the background difference method is D(x, y), the output of the mixed Gaussian method is G(x, y), and the output of the improved foreground detection method is R(x, y). ),but

Then use some binarization methods to process the picture, and you can get a clear binarized image of the target person.

After foreground extraction, open and close the image and select the largest connected domain to roughly determine the position of the person's foreground, and intercept the corresponding position to obtain the corresponding RGB image of the person's foreground.

2. Model selection

Adopt the champion model ResNet of 2015 as the network model of the present invention. The ResNet network can effectively solve the problem that as the network depth increases, the accuracy of the algorithm tends to saturate and decline rapidly. At the same time, its parameter amount is lower than that of VGGNet, and the effect is very significant. While greatly improving the accuracy of the model, the training speed has also been greatly improved. This is mainly due to the construction idea of the residual module it uses, as shown in Figure 2.

The following table shows the comparison of ResNet-34 and VGG-16 in ImageNet2012

network RESNET-34 VGG-16 Calculations 3.6 billion FLOPs 15.3 billion FLOPs Accuracy top-1 0.733 0.715

The network architecture of ResNet is shown in the following table:

pre-trained model

Before officially using the preprocessed images for model training, ImageNet is used to pre-train the parameters of the convolutional neural network. Professor Bengio and others pointed out that the method of random initialization of the model often makes the probability of the model entering the local minimum is very high, and the use of pre-training can make the model more effective. In actual operation, the function of the pre-training model is to initialize the parameters of the network to the model parameters trained by ImageNet on the network. However, because ImageNet is divided into 1000 categories, and the present invention only needs to divide images into two categories, it is necessary to modify the fully connected layer, change the value of num_output from the original 1000 to 2, and modify the name of the fully connected layer .

model training

The training process of the network is completed based on the Caffe platform. Caffe is built according to a hypothesis of the neural network: all calculations are represented in the form of layers, and the effect of the layers is to output the calculated results according to the input data. As far as convolution is concerned, if the input is an image, then perform a convolution operation with the parameters of this layer, and finally output the result of the convolution. Each layer requires two operations: 1) the forward path, which calculates the output data from the input data; 2) the reverse path, which calculates the gradient relative to the input based on the above gradient value. When each layer has completed these two functions, many layers can be connected into a network. The function of the network is to calculate the desired output based on the input data (image, voice or other information forms). During the training process, the loss function and gradient of the output model can be calculated according to the known label results, and then the parameters of the network can be further updated according to the gradient value.

Before training the model, it is necessary to encapsulate the image database, that is, the image and the corresponding label, that is, to encapsulate the image set into the form of a database: Lmdb and Leveldb. Specifically, the convert_imageset command is called to convert the data format. The construction of the ResNet34 model is realized by defining trian_val.prototxt, which defines the specific network structure of ResNet. The organizational structure of this file is in the form of a class structure. Each layer structure includes many parameters. For example, the bottom parameter indicates the bottom layer input, the top parameter indicates the result output to the next layer, and the param parameter indicates the parameters of this layer, including num_output indicating the number of filters, kernel_size indicating the size of the filter, and stride indicating the step size . ResNet's training model is written through the file solver.prototxt, and each line in this file represents a training parameter. Among them, some common training parameters, such as the net parameter, are used to specify the model used, while the max_iter parameter indicates the maximum number of iterations set, and the snapshot_prefix parameter indicates its prefix name when the model is saved, etc.

At the same time, the loss function curve during model training can also be drawn on the Matlab platform. The loss value is plotted every iteration, and the accuracy is plotted every 100 iterations. As shown in Figure 3.

model testing

Use the trained model to classify a single picture and output the classification result, this step is also implemented on Matlab. Refer to Figure 4 for the specific process.

When obtaining the forward propagation results, you only need to call the function net.forward that comes with matlab to obtain the probability of the image under each label. Data exists in three dimensions at each convolutional layer. It can be seen as a stack of many two-dimensional images, each of which is called a feature map. If the input layer is a grayscale image, then there is only one feature map; if the input layer is a color image, there are generally three feature maps (red, green, and blue). There are many convolution kernels (kernels) between layers, and each feature map of the previous layer is convolved with each convolution kernel to generate a feature map of the next layer.

1. First use the improved foreground detection method to extract the characters, and the effect diagrams of the background difference method and the mixed Gaussian model to extract the foreground are shown in Figure 5 and Figure 6:

Using the improved foreground detection method, the effect diagram after "ANDing" the results of the background difference method and the mixed Gaussian model is shown in Figure 7.

Then, perform simple binarization processing and extraction of the largest connected domain to determine the range of the foreground of the person, so as to intercept the RGB image corresponding to the person on the original image

2. Perform model training and performance testing on the public dataset UR Fall Detection Dataset. Among them, the training set has 7381 pictures, and the test set has 1326 pictures. All the pictures in the data set are firstly subjected to the foreground extraction operation, and the pictures shown in Figure 8 are obtained and then put into the network for model training.

3. On the ImageNet data set, the ResNet network used in the present invention is pre-trained to obtain a pre-trained model. The pictures after preprocessing (foreground detection, whitening, normalization) of the training set are input into the network for training.

4. Use the test set to test the accuracy of the trained model, test the loss value every iteration, test the accuracy every 100 iterations, and the final convergence accuracy is 96.7%

5. Use the Matlab platform to test the output of a single image, and realize the visualization of the convolution kernel and feature map. The visualization picture of the convolution kernel is shown in Figure 9

Call the function feature_map that comes with matlab to realize the visualization of the feature map, as shown in Figure 10 and Figure 11.

It can be seen from Figure 10 and Figure 11 that the underlying convolution kernel is mainly used to extract basic features such as the outline of a person.

Both the training set and the test set of the present invention come from the UR Fall Detection Dataset. The dataset is used for model training and accuracy testing on the caffe platform, and is accelerated by cuDNN. The final display accuracy reaches 96.7%, and the time complexity is 49ms .

The various technical features of the above-mentioned embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the various technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, should be considered as within the scope of this specification.

The above-mentioned embodiments only express several implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but should not be construed as limiting the patent scope of the invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.

Claims (10)

1. A fall detection method based on convolutional neural network, characterized in that, comprising: Training the convolutional neural network, the training convolutional neural network specifically includes: Perform preprocessing on each frame of image obtained, and the preprocessing work includes foreground extraction, normalization, and whitening operations in sequence; Pre-train the ResNet network on the ImageNet dataset to obtain a pre-trained model; The step "preprocess each frame of image obtained, the preprocessing work includes foreground extraction, normalization, and whitening operations in turn;" the processed picture is put into the pre-training model for model training, and the obtained the parameters of the model; and Input the test set into the trained model, and use the test set to test the accuracy of the model; Use the trained convolutional neural network to detect pictures; Wherein, the method for foreground extraction specifically includes: Image processing using background subtraction; Image processing using a mixture of Gaussian models; Sum the result of processing the image using the background subtraction method and the result of processing the image using the mixture Gaussian model. 2. the fall detection method based on convolutional neural network according to claim 1, is characterized in that, in step " each frame image that obtains is carried out preprocessing, the work of preprocessing comprises foreground extraction and normalization successively Whitening and whitening operations; ", each frame of image acquired is obtained by reading the video file. 3. The fall detection method based on convolutional neural network according to claim 1, characterized in that, after the step "use the trained convolutional neural network to detect pictures", the training convolutional neural network specifically also includes : Display the detection effect diagram of each frame of the picture, and realize the visualization of the convolution kernel of the model. 4. the fall detection method based on convolutional neural network according to claim 3, is characterized in that, step " shows the detection effect figure of each frame picture, and realizes the convolution kernel visualization of model " on the matlab platform Display the detection effect diagram of each frame of the picture, and realize the visualization of the convolution kernel of the model. 5. the fall detection method based on convolutional neural network according to claim 1, is characterized in that, step " utilizes background difference method to process image; " specifically comprises: Take the average value of the previous few frame images and use it as the initial background image B _t ; The grayscale of the previous frame image and the background image is subtracted, and its absolute value is N _t (x, y) The formula is N _t (x,y)＝|I _t (x,y)-B _t (x,y)| For the pixel (x, y) of the current frame, if there is |I _t (x, y)-B _t (x, y)|≥T, then the pixel point is the foreground point, that is, the current image frame is updated as; Update the background image with the current frame image. 6. the fall detection method based on convolutional neural network according to claim 1, is characterized in that, step " utilizes mixture Gaussian model processing image; " specifically comprises: When the Gaussian mixture model is used to model the background, the pixel value of each pixel of the sequence image can be simulated by k Gaussian models, so at time t, the probability density function of a certain pixel value can be expressed as: Among them, w _{i, t} represents the weight of the Gaussian model, and the probability density function of the Gaussian model is expressed as: Next, the K Gaussian mixture models are sorted according to the size of the quotient of the weight divided by the standard deviation, and then the previous B Gaussian models are selected for discrimination, where the value of B is expressed as: Put each pixel in the new image frame into the sorted K Gaussian models for judgment, and the judgment conditions are: ||X _t -μ _t ||≤β∑ ^1/2 Among the previous B Gaussian models, if one of the Gaussian models meets the above conditions, then the pixel is judged as the background, and if none of the above conditions is satisfied in the B Gaussian models, the pixel is judged to belong to the foreground. For each Gaussian model, assuming that the above conditions are not true, the weight of this Gaussian model must be reduced. If the above conditions are true, the Gaussian mixture model must be updated. The specific operation method is shown in the following formula: w _i,t = (1-λ)w _i,t-1 +λBM _t μ _i,t ＝(1-α)μ _i,t-1 +αX _i,t ∑ _i,t ＝(1-α)∑ _i,t-1 +α(X _i,t -μ _i,t )(X _i,t -μ _i,t ) ^T α=λ/w _i,t Among them, if the pixel is the foreground, BM=0, otherwise BM=1, and finally an initial Gaussian model is used to replace the Gaussian model with the smallest weight, the threshold T, the learning rate λ, and the parameter β are all constants specified in advance. 7. the fall detection method based on convolutional neural network according to claim 1, is characterized in that, In the step "input the test set into the trained model, and use the test set to test the accuracy of the model;", the test set comes from the UR Fall Detection Dataset. 8. A computer device, comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, characterized in that, when the processor executes the program, any one of claims 1 to 7 is realized The steps of the method. 9. A computer-readable storage medium, on which a computer program is stored, wherein, when the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are realized. 10. A processor, wherein the processor is used to run a program, wherein the method according to any one of claims 1 to 7 is executed when the program is running.