System and method for monitoring parking lots based on remote seats

Abstract

The present invention proposes a system and method for remote agent supervision of parking lots, aiming to improve the efficiency and safety of parking lot management through remote monitoring and automation technology. The system mainly includes image acquisition equipment, a remote agent center, a background subtraction module, a motion feature extraction module and anomaly detection module. The data processing module uses background subtraction and optical flow methods to pre-process image data to extract foreground and motion features, and combines anomaly detection algorithms to detect abnormal behavior and fire. The remote agent center receives and analyzes the processed data through the network, and combines deep learning algorithms to monitor and warn abnormal behaviors in real time, while providing remote emergency response suggestions to facilitate managers to take timely action. Through deep learning technology, the system achieves efficient abnormal behavior detection and emergency response, improving the safety and management efficiency of parking lots.

Description

Technical Field

[0001] The present invention provides a parking lot supervision method, and in particular relates to a system and method for supervising a parking lot based on a remote seat. Background Art

[0002] With the acceleration of urbanization and the rapid growth of the number of motor vehicles, parking difficulties are becoming increasingly prominent and have become a major challenge for urban traffic management. Remote agent monitoring technology enables real-time monitoring and management of parking lot operations through a remote monitoring and management center. The core of this technology is the use of sensors, cameras, data transmission networks, and other equipment to transmit real-time parking lot data to a remote monitoring center for real-time monitoring and management by agents. Remote agent monitoring technology can significantly improve the efficiency and accuracy of parking lot management, reduce manual intervention, and enhance the user experience.

[0003] Deep learning and image processing technologies are already widely used in intelligent parking systems. By capturing real-time images of parking lots through cameras and applying deep learning algorithms to image processing and analysis, they enable functions such as vehicle identification, license plate recognition, and parking space detection. These technologies not only enhance parking automation but also provide accurate, real-time data support for remote monitoring.

[0004] While intelligent parking systems have made significant progress in some areas, existing technologies still have some shortcomings. For example, traditional abnormal behavior detection methods are prone to false positives and false negatives in complex scenarios. Existing remote agent monitoring systems still need to improve the real-time and reliability of data transmission. These issues urgently require technical personnel to address. Summary of the Invention

[0005] The present invention aims to at least solve the technical problems existing in the prior art, and in particular innovatively proposes a system and method for remote parking supervision.

[0006] In order to achieve the above-mentioned object of the present invention, the present invention provides a system for remotely monitoring a parking lot, comprising:

[0007] Cloud computing platform: uploads image data obtained by image acquisition devices to the cloud for deep learning model training and more complex analysis.

[0008] First, abnormal behavior judgment is performed. Before abnormal behavior judgment, the background image needs to be subtracted;

[0009] Background subtraction is used to remove static backgrounds in videos and highlight moving objects. An adaptive Gaussian mixture model is used to identify the grayscale value of the background for subsequent background subtraction.

[0010]

[0011] Among them: B t+1 (x,y) is the background value at position (x,y) at time t+1; α t and β t is the dynamic learning rate, subscript t is the current time, which determines the update weight of the image grayscale value at time t and the background model;

[0012] I t (x,y) is the grayscale value of the pixel at position (x,y) at time t;

[0013] B t (x+i, y+j) is the grayscale value of the background model at position (x+i, y+j); i is the movement value of the image pixel's horizontal coordinate x, and j is the movement value of the image pixel's vertical coordinate y;

[0014] w ij is the Gaussian weight, and m is the window size, which is used to set according to the background object.

[0015] Dynamic learning rate α at time t t , dynamic learning rate β for background position t And the background position Gaussian weight w ij The calculation method is:

[0016]

[0017]

[0018] Among them, B t (x,y) is the background value at position (x,y) at time t,

[0019] The initial learning rate α0 at time t and the initial learning rate β0 at the background position, and α0+β0=1, M is the total number of pixels in the image, which is used for normalization after difference summation,

[0020] Represents pixel I at time t t (x, y) and the background value B at time t t The average of the global differences between (x,y),

[0021] is the local variance term, is the variance of the image grayscale value of the background model at position (x, y),

[0022] Represents time t I t (x, y) and the image grayscale value B of the background model tThe difference between (x,y) in the local binary pattern LBP, LBP is a texture descriptor used to capture local texture features,

[0023] λ is a parameter that controls the change of weight,

[0024] Index term Represents the similarity between the pixel at position (x+i,y+j) and the position (x,y). The similarity is measured by the square of the difference between the two, then multiplied by negative λ and taking the exponential.

[0025] Normalization term The similarity indices of all neighborhood pixels are summed up to normalize the weights to ensure that the sum of all weights is 1.

[0026] The background subtraction module also includes using the foreground mask to subtract the redundant static content of the background.

[0027]

[0028] in,

[0029] I t (x,y) represents the grayscale value of the pixel at position (x,y) at time t,

[0030] B t (x,y) represents the background value at position (x,y) at time t,

[0031] Threshold: A preset threshold used to determine the current pixel gray value I t (x, y) and the image grayscale value B of the background model t Is the difference between (x,y) significant?

[0032] F t (x,y) is the foreground mask, F t (x,y)=1 means that at the position (x,y) at time t, the image gray value I t (x, y) and the image grayscale value B of the background model t If the difference between (x, y) is greater than the preset threshold, the pixel is considered to belong to the foreground, that is, there is a moving object or change at the position (x, y) at time t.

[0033] F t (x,y)=0 means that at the position (x,y) at time t, the image gray value I t (x, y) and the image grayscale value B of the background model t If the difference between (x, y) is less than or equal to the preset threshold, the pixel is considered to belong to the background, that is, there is no significant change at the position (x, y) at time t.

[0034] Feature extraction:

[0035] After background subtraction is completed, feature extraction begins to analyze the behavioral content of parking images in the parking lot.

[0036] In the foreground mask F t In the area where (x,y)=1, the optical flow field is used to describe the changes of image pixels over time:

[0037] I F (x,y,t) is the foreground mask F t In the area where (x, y) = 1, the grayscale value of the pixel image at position (x, y) at time t;

[0038] Through Taylor expansion, we can get

[0039]

[0040] and is the spatial gradient of the image grayscale value, is the time gradient of the image gray value;

[0041] (x,y) is the image coordinate, t is the time, u and v are the components of the optical flow in the x and y directions respectively;

[0042] Calculate the magnitude of the motion vector:

[0043]

[0044] Calculate the direction of the motion vector:

[0045]

[0046] For the calculated motion features magnitude and θ, the feature vector is extracted as f i ,

[0047] f i =(magnitude,θ)

[0048] The probability distribution of normal behavior can be expressed as

[0049]

[0050] in, is the kth Gaussian distribution, calculated as:

[0051]

[0052] μ kis the mean of each motion vector magnitude and each motion vector direction θ, ∑k is the covariance matrix of each motion vector magnitude and each motion vector direction θ, is the inverse matrix of the covariance matrix ∑k, K is the total number of Gaussian distributions, T represents the transpose; f i is the feature vector of the i-th input;

[0053] π k is the weight of the k-th Gaussian distribution;

[0054] Gaussian distribution weight π k The update method is:

[0055]

[0056] is the weight of the k-th Gaussian distribution in the t-th iteration, is the weight of the k-th Gaussian distribution in the t+1-th iteration, is the expected value of the number of data points generated by the k-th Gaussian distribution in the t-th iteration, γ is a dynamic adjustment parameter, δ is a regularization parameter, K is the total number of Gaussian distributions, and η is the learning rate used to adjust the step size of gradient descent. is the gradient of the loss function with respect to the weights, N represents the total number of data points, and in each iteration, the number of all data points remains unchanged.

[0057] The anomaly detection module includes using self-supervised learning methods to train the model so that it can identify unseen abnormal behaviors. The model based on AutoEncoder can learn the characteristics of normal behavior. When abnormal behavior is detected, there will be a higher reconstruction error, which is calculated as follows:

[0058] Encoder:

[0059] h=f(W e f i +b e )

[0060] Decoder:

[0061]

[0062] Hybrid reconstruction error:

[0063]

[0064] Among them, f i is the feature vector of the i-th input, h is the encoded feature, is the reconstructed feature vector after the i-th decoding, and n is the dimension of the feature vector;

[0065] f() and g() are the function representations of the encoder and decoder respectively, W e and W d are the weight matrices of the encoder and decoder, b e and b d are the bias vectors of the encoder and decoder respectively;

[0066] is the mean square error term, is the log-likelihood term,

[0067] is the weight parameter used to adjust the weight between the mean square error term and the log likelihood term, σ 2 is the variance of the Gaussian distribution;

[0068] L e Is the reconstruction error, which is used to measure the difference between the input data and the reconstructed data. The reconstruction error of the new feature vector is calculated in the inference phase. If the isomorphism error is greater than the threshold τ, it is judged as an anomaly.

[0069] When the isomorphism error L e When τ > τ, the system will be judged as abnormal behavior. The setting of the threshold τ is determined by the mean and variance of the reconstruction error:

[0070]

[0071] in,

[0072] is the reconstruction error L e The mean of

[0073] is the reconstruction error L e The standard deviation of

[0074] χ is a constant used to adjust the sensitivity of the threshold;

[0075] About the reconstruction error L e The mean and standard deviation The calculation methods are:

[0076]

[0077] in, is the i-th reconstruction error, P represents the number of all samples involved in calculating the reconstruction error;

[0078] When the image recognition system detects abnormal behavior, it will provide an alarm signal to the abnormal behavior subject and the staff of the remote monitoring center to facilitate the response to the abnormality.

[0079] Responding to abnormal behavior:

[0080] Real-time alert system: When abnormal behavior is detected, the edge device immediately triggers an alert and notifies the monitoring center through the network.

[0081] The system can automatically execute some preset emergency measures, such as closing gates and turning on warning lights.

[0082] Deep learning-assisted decision-making: Deep learning models are used to analyze historical emergency data and optimize emergency response plans. These models can predict the effects of different emergency measures and help develop optimal emergency response strategies.

[0083] Dynamic emergency response plan: Dynamically adjust the emergency response plan based on real-time monitoring data and prediction results of deep learning models to improve the flexibility and effectiveness of emergency response.

[0084] For example, after the model detects a fire hazard, it notifies remote staff to respond. Different levels of emergency needs correspond to different degrees of alarms and emergency response measures.

[0085] Intelligent customer service system: When abnormal behavior is detected, the system can generate automated notifications through NLP technology and communicate with users through voice or text to provide guidance and assistance.

[0086] Emergency response team collaboration: Through the cloud platform, abnormal behavior detection results and emergency response suggestions are shared in real time to ensure that the emergency response team can quickly collaborate to handle emergencies.

[0087] In summary, due to the adoption of the above technical solution, the intelligent parking system based on remote seat supervision of the present invention, combined with deep learning technology, realizes efficient and accurate parking lot management, which has important practical significance and application value.

[0088] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0090] Figure 1 It is an overall schematic diagram of the present invention. DETAILED DESCRIPTION

[0091] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0092] The present invention provides a method for supervising a parking lot based on a remote seat. Figure 1 As shown, including the following:

[0093] Cloud computing platform: uploads image data obtained by image acquisition devices to the cloud for deep learning model training and more complex analysis.

[0094] First, abnormal behavior judgment is performed. Before abnormal behavior judgment, the background image needs to be subtracted;

[0095] Background subtraction is used to remove static backgrounds in videos and highlight moving objects. An adaptive Gaussian mixture model is used to identify the grayscale value of the background for subsequent background subtraction.

[0096]

[0097] Among them: B t+1 (x,y) is the background value at position (x,y) at time t+1; α t and β t is the dynamic learning rate, subscript t is the current time, which determines the update weight of the image grayscale value at time t and the background model;

[0098] I t (x,y) is the grayscale value of the pixel at position (x,y) at time t;

[0099] B t (x+i, y+j) is the grayscale value of the background model at position (x+i, y+j); i is the movement value of the image pixel's horizontal coordinate x, and j is the movement value of the image pixel's vertical coordinate y;

[0100] w ij is the Gaussian weight, and m is the window size, which is used to set according to the background object.

[0101] Dynamic learning rate α at time t t , dynamic learning rate β for background position t And the background position Gaussian weight w ij The calculation method is:

[0102]

[0103] Among them, B t(x,y) is the background value at position (x,y) at time t,

[0104] The initial learning rate α0 at time t and the initial learning rate β0 at the background position, and α0+β0=1, M is the total number of pixels in the image, which is used for normalization after difference summation,

[0105] Represents pixel I at time t t (x, y) and background value B at time t t The average of the global differences between (x,y),

[0106] is the local variance term, is the variance of the image grayscale value of the background model at position (x, y),

[0107] Represents time t I t (x, y) and the image grayscale value B of the background model t The difference between (x,y) in the local binary pattern LBP, LBP is a texture descriptor used to capture local texture features,

[0108] λ is a parameter that controls the change of weight,

[0109] Index term Represents the similarity between the pixel at position (x+i,y+j) and the position (x,y). The similarity is measured by the square of the difference between the two, then multiplied by negative λ and taking the exponential.

[0110] Normalization term The similarity indices of all neighborhood pixels are summed up to normalize the weights to ensure that the sum of all weights is 1.

[0111] Use the foreground mask to remove the redundant static content of the background.

[0112]

[0113] in,

[0114] I t (x,y) represents the grayscale value of the pixel at position (x,y) at time t,

[0115] B t (x,y) represents the background value at position (x,y) at time t,

[0116] Threshold: A preset threshold used to determine the current pixel gray value I t (x, y) and the image grayscale value B of the background model t Is the difference between (x,y) significant?

[0117] F t (x,y) is the foreground mask, F t (x,y)=1 means that at the position (x,y) at time t, the image gray value I t (x, y) and the image grayscale value B of the background model t If the difference between (x, y) is greater than the preset threshold, the pixel is considered to belong to the foreground, that is, there is a moving object or change at the position (x, y) at time t.

[0118] F t (x,y)=0 means that at the position (x,y) at time t, the image gray value I t (x, y) and the image grayscale value B of the background model t If the difference between (x, y) is less than or equal to the preset threshold, the pixel is considered to belong to the background, that is, there is no significant change at the position (x, y) at time t.

[0119] For example, the gray value of the image at time t is I t (x,y) is

[0120]

[0121] The image gray value B of the background model t (x,y) is

[0122]

[0123] Assume that the threshold at the current time t is 5 and the foreground mask F t The result of calculating (x,y) is

[0124]

[0125] Feature extraction:

[0126] After background subtraction is completed, feature extraction begins to analyze the behavioral content of parking images in the parking lot.

[0127] In the foreground mask F t In the area where (x,y)=1, the optical flow field is used to describe the changes of image pixels over time:

[0128] I F (x,y,t) is the foreground mask F t In the area where (x, y) = 1, the grayscale value of the pixel image at the position (x, y) at time t

[0129] Through Taylor expansion, we can get

[0130]

[0131] and is the spatial gradient of the image grayscale value, is the time gradient of the image gray value;

[0132] (x,y) is the image coordinate, t is the time, u and v are the components of the optical flow in the x and y directions respectively;

[0133] Calculate the magnitude of the motion vector:

[0134]

[0135] Calculate the direction of the motion vector:

[0136]

[0137] For the calculated motion features magnitude and θ, the feature vector is extracted as f i ,

[0138] f i =(magnitude,θ)

[0139] The probability distribution of normal behavior can be expressed as:

[0140]

[0141] in, is the kth Gaussian distribution, calculated as,

[0142]

[0143] μ k is the mean of each motion vector magnitude and each motion vector direction θ, ∑k is the covariance matrix of each motion vector magnitude and each motion vector direction θ, is the inverse matrix of the covariance matrix ∑k, K is the total number of Gaussian distributions, T represents the transpose; f i is the feature vector of the i-th input;

[0144] π k is the weight of the k-th Gaussian distribution;

[0145] Gaussian distribution weight π k The update method is:

[0146]

[0147] is the weight of the k-th Gaussian distribution in the t-th iteration, is the weight of the k-th Gaussian distribution in the t+1-th iteration, is the expected value of the number of data points generated by the k-th Gaussian distribution in the t-th iteration, γ is a dynamic adjustment parameter, δ is a regularization parameter, K is the total number of Gaussian distributions, and η is the learning rate used to adjust the step size of gradient descent. is the gradient of the loss function with respect to the weights, N represents the total number of data points, and in each iteration, the number of all data points remains unchanged.

[0148] The anomaly detection module includes using self-supervised learning methods to train the model so that it can identify unseen abnormal behaviors. The model based on AutoEncoder can learn the characteristics of normal behavior. When abnormal behavior is detected, there will be a higher reconstruction error, which is calculated as follows:

[0149] Encoder:

[0150] h=f(W e f i +b e )

[0151] Decoder:

[0152]

[0153] Hybrid reconstruction error:

[0154]

[0155] Among them, f i is the feature vector of the i-th input, h is the encoded feature, is the reconstructed feature vector after the i-th decoding, and n is the dimension of the feature vector;

[0156] f() and g() are the function representations of the encoder and decoder respectively, W e and W d are the weight matrices of the encoder and decoder, b e and b d are the bias vectors of the encoder and decoder respectively;

[0157] is the mean square error term, is the log-likelihood term,

[0158] is the weight parameter used to adjust the weight between the mean square error term and the log likelihood term, σ 2 is the variance of the Gaussian distribution;

[0159] L eIs the reconstruction error, which is used to measure the difference between the input data and the reconstructed data. The reconstruction error of the new feature vector is calculated in the inference phase. If the isomorphism error is greater than the threshold τ, it is judged as an anomaly.

[0160] When the isomorphism error L e When τ > τ, the system will be judged as abnormal behavior. The setting of the threshold τ is determined by the mean and variance of the reconstruction error:

[0161]

[0162] in,

[0163] is the reconstruction error L e The mean of

[0164] is the reconstruction error L e The standard deviation of

[0165] χ is a constant used to adjust the sensitivity of the threshold;

[0166] About the reconstruction error L e The mean and standard deviation The calculation methods are:

[0167]

[0168] in, is the i-th reconstruction error, P represents the number of all samples involved in calculating the reconstruction error;

[0169] When the image recognition system detects abnormal behavior, it will provide an alarm signal to the abnormal behavior subject and the staff of the remote monitoring center to facilitate the response to the abnormality.

[0170] Responding to abnormal behavior:

[0171] Real-time alert system: When abnormal behavior is detected, the edge device immediately triggers an alert and notifies the monitoring center through the network.

[0172] The system can automatically execute some preset emergency measures, such as closing gates and turning on warning lights.

[0173] Deep learning-assisted decision-making: Deep learning models are used to analyze historical emergency data and optimize emergency response plans. These models can predict the effects of different emergency measures and help develop optimal emergency response strategies.

[0174] Dynamic emergency response plan: Dynamically adjust the emergency response plan based on real-time monitoring data and prediction results of deep learning models to improve the flexibility and effectiveness of emergency response.

[0175] For example, after the model detects a fire hazard, it notifies remote staff to respond. Different levels of emergency needs correspond to different degrees of alarms and emergency response measures.

[0176] Intelligent customer service system: When abnormal behavior is detected, the system can generate automated notifications through NLP technology and communicate with users through voice or text to provide guidance and assistance.

[0177] Emergency response team collaboration: Through the cloud platform, abnormal behavior detection results and emergency response suggestions are shared in real time to ensure that the emergency response team can quickly collaborate to handle emergencies.

[0178] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A system based on remote seat supervision of parking lot, characterized in that: include: The background subtraction module uses the background subtraction algorithm to separate the background from the foreground of the image data obtained by the image acquisition device; The background subtraction module includes: Use the adaptive Gaussian mixture model to identify the grayscale value of the background for subsequent background subtraction. Among them: B t+1 (x,y) is the background value at position (x,y) at time t+1; α t and β t is the dynamic learning rate, subscript t is the current time, which determines the update weight of the image grayscale value at time t and the background model; I t (x,y) is the grayscale value of the pixel at position (x,y) at time t; B t (x+i, y+j) is the grayscale value of the background model at position (x+i, y+j); i is the movement value of the image pixel's horizontal coordinate x, and j is the movement value of the image pixel's vertical coordinate y; w ij is the Gaussian weight, m ​​is the window size, which is used to set according to the background object; The dynamic learning rate α at time t t , dynamic learning rate β for background position t And the background position Gaussian weight w ij include: Dynamic learning rate α at time t t , dynamic learning rate β for background position t And the background position Gaussian weight w ij The calculation method is: Among them, B t (x,y) is the background value at position (x,y) at time t, The initial learning rate α0 at time t and the initial learning rate β0 at the background position, and α0+β0=1, M is the total number of pixels in the image, which is used for normalization after difference summation, Represents pixel I at time t t (x, y) and background value B at time t t The average of the global differences between (x,y), is the local variance term, is the variance of the image grayscale value of the background model at position (x, y), Represents time t I t (x, y) and the image grayscale value B of the background model t The difference between (x,y) in the local binary pattern LBP, LBP is a texture descriptor used to capture local texture features, λ is a parameter that controls the change of weight, Index term Represents the similarity between the pixel at position (x+i,y+j) and the position (x,y). The similarity is measured by the square of the difference between the two, then multiplied by negative λ and taking the exponential. Normalization term The similarity index of all neighboring pixels is summed to normalize the weights to ensure that the sum of all weights is 1; The motion feature extraction module uses the optical flow method to calculate the motion vector in the foreground area and extract the motion features; The anomaly detection module uses a self-supervised learning method to train the model and uses the isomorphism error to measure the difference between the input data and the reconstructed data to determine whether it is an anomaly.

2. The system for remote parking lot supervision according to claim 1 is characterized in that: The background subtraction module includes: Use the foreground mask to remove the redundant static content of the background. in, I t (x,y) represents the grayscale value of the pixel at position (x,y) at time t, B t (x,y) represents the background value at position (x,y) at time t, Threshold represents a preset threshold value, which is used to judge the current pixel gray value I t (x, y) and the image grayscale value B of the background model t Is the difference between (x,y) significant? F t (x,y) is the foreground mask, F t (x,y)=1 means that at the position (x,y) at time t, the image gray value I t (x, y) and the image grayscale value B of the background model t If the difference between (x, y) is greater than the preset threshold, the pixel is considered to belong to the foreground, that is, there is a moving object or change at the position (x, y) at time t; F t (x,y)=0 means that at the position (x,y) at time t, the image gray value I t (x, y) and the image grayscale value B of the background model t If the difference between (x, y) is less than or equal to the preset threshold, the pixel is considered to belong to the background, that is, there is no change at the position (x, y) at time t.

3. The system for remote parking lot supervision according to claim 1 is characterized in that: The motion feature extraction module includes: In the foreground mask F t In the area where (x,y)=1, the optical flow field is used to describe the changes of image pixels over time: I F (x,y,t) is the foreground mask F t In the area where (x, y) = 1, the grayscale value of the pixel image at position (x, y) at time t; By Taylor expansion, we get: and isI F (x,y,t) in the foreground mask F t In the region where (x, y) = 1, the spatial gradient of the grayscale value of the pixel image at the position (x, y) at time t is isI F (x,y,t) in the foreground mask F t In the region where (x, y) = 1, the temporal gradient of the grayscale value of the pixel image at position (x, y) at time t; (x,y) is the image coordinate, t is the time, u and v are the components of the optical flow in the x and y directions respectively; Calculate the magnitude of the motion vector, Calculate the direction of the motion vector, For the calculated motion features magnitude and θ, the feature vector is extracted as f i , f i =(magnitude,θ) The probability distribution of normal behavior is expressed as, in, is the kth Gaussian distribution, calculated as, μ k is the mean of each motion vector magnitude and each motion vector direction θ, ∑k is the covariance matrix of each motion vector magnitude and each motion vector direction θ, is the inverse matrix of the covariance matrix ∑k, K is the total number of Gaussian distributions, T represents the transpose; f i is the feature vector of the i-th input; π k is the weight of the k-th Gaussian distribution; Gaussian distribution weight π k The update method is: is the weight of the k-th Gaussian distribution in the t-th iteration, is the weight of the k-th Gaussian distribution in the t+1-th iteration, is the expected value of the number of data points generated by the k-th Gaussian distribution in the t-th iteration, γ is a dynamic adjustment parameter, δ is a regularization parameter, K is the total number of Gaussian distributions, and η is the learning rate used to adjust the step size of gradient descent. is the gradient of the loss function with respect to the weights, N represents the total number of data points, and in each iteration, the number of all data points remains unchanged.

4. The system for remote parking lot monitoring according to claim 1 is characterized in that: The anomaly detection module includes: The model is trained using self-supervised learning methods to enable it to identify abnormal behavior. The AutoEncoder-based model learns the characteristics of normal behavior. When abnormal behavior is detected, there will be reconstruction error, which is calculated as follows: Encoder: h=f(W e f i +b e ) Decoder: Hybrid reconstruction error: Among them, f i is the feature vector of the i-th input, h is the encoded feature, is the reconstructed feature vector after the i-th decoding, and n is the dimension of the feature vector; f() and g() are the function representations of the encoder and decoder respectively, W e and W d are the weight matrices of the encoder and decoder, b e and b d are the bias vectors of the encoder and decoder respectively; is the mean square error term, is the log-likelihood term, is the weight parameter used to adjust the weight between the mean square error term and the log likelihood term, σ 2 is the variance of the Gaussian distribution; L e Is the reconstruction error, which is used to measure the difference between the input data and the reconstructed data. The reconstruction error of the new feature vector is calculated in the inference phase. If the isomorphism error is greater than the threshold τ, it is judged as an anomaly.

5. The system for remote parking lot supervision according to claim 4 is characterized in that: The calculation method of the threshold τ includes: When the isomorphism error L e When τ > τ, the system will be judged as abnormal behavior. The setting of the threshold τ is determined by the mean and variance of the reconstruction error: in, is the reconstruction error L e The mean of is the reconstruction error L e The standard deviation of χ is a constant used to adjust the sensitivity of the threshold; About the reconstruction error L e The mean and standard deviation The calculation methods are: in, is the i-th reconstruction error, P represents the number of all samples involved in calculating the reconstruction error; When the image recognition system detects abnormal behavior, it will provide an alarm signal to the abnormal behavior subject and the staff of the remote monitoring center to facilitate the response to the abnormality.

6. A method for monitoring a parking lot based on a remote seat, characterized in that: The following steps are involved: S1, the image data acquired by the image acquisition device is separated from the background and foreground by a background subtraction algorithm; The background subtraction algorithm includes: Use the adaptive Gaussian mixture model to identify the grayscale value of the background for subsequent background subtraction. Among them: B t+1 (x,y) is the background value at position (x,y) at time t+1; α t and β t is the dynamic learning rate, subscript t is the current time, which determines the update weight of the image grayscale value at time t and the background model; I t (x,y) is the grayscale value of the pixel at position (x,y) at time t; B t (x+i, y+j) is the grayscale value of the background model at position (x+i, y+j); i is the movement value of the image pixel's horizontal coordinate x, and j is the movement value of the image pixel's vertical coordinate y; w ij is the Gaussian weight, m ​​is the window size, which is used to set according to the background object; The dynamic learning rate α at time t t , dynamic learning rate β for background position t And the background position Gaussian weight w ij include: Dynamic learning rate α at time t t , dynamic learning rate β for background position t And the background position Gaussian weight w ij The calculation method is: Among them, B t (x,y) is the background value at position (x,y) at time t, The initial learning rate α0 at time t and the initial learning rate β0 at the background position, and α0+β0=1, M is the total number of pixels in the image, which is used for normalization after difference summation, Represents pixel I at time t t (x, y) and background value B at time t t The average of the global differences between (x,y), is the local variance term, is the variance of the image grayscale value of the background model at position (x, y), Represents time t I t (x, y) and the image grayscale value B of the background model t The difference between (x,y) in the local binary pattern LBP, LBP is a texture descriptor used to capture local texture features, λ is a parameter that controls the change of weight, Index term Represents the similarity between the pixel at position (x+i,y+j) and the position (x,y). The similarity is measured by the square of the difference between the two, then multiplied by negative λ and taking the exponential. Normalization term The similarity index of all neighboring pixels is summed to normalize the weights to ensure that the sum of all weights is 1; S2, uses the optical flow method to calculate the motion vector in the foreground area and extract motion features; S3 uses self-supervised learning methods to train the model and uses isomorphism error to measure the difference between the input data and the reconstructed data to determine whether it is an anomaly.

Images