A heavy-load traffic abnormal event detection and state judgment system

Abstract

A heavy-load traffic abnormal event detection and state discrimination system belongs to the field of intelligent transportation, including a data acquisition and annotation module, an image processing module, a traffic state analysis module, a state judgment and abnormal event recognition module and a transfer learning module; the data acquisition and annotation module includes a data acquisition unit and a data annotation unit; the image processing module includes an image enhancement unit and an image compression and reconstruction unit; the traffic state analysis module includes a traffic bottleneck recognition unit and a traffic wave analysis unit; the state judgment and abnormal event recognition module includes a vehicle statistics unit, a congestion state recognition unit, a state transition analysis unit and a compressed sensing model establishment unit; the transfer learning module includes a transfer learning unit. The present invention can reduce the false detection rate of abnormal traffic flow events under bad weather conditions, so that abnormal traffic events can be detected around the clock, thereby reducing the probability of traffic events under bad weather conditions.

Description

Technical Field

[0001] The invention belongs to the field of intelligent transportation, and in particular to a heavy-load traffic abnormal event detection and state discrimination system. Background Art

[0002] Traffic event perception is a monitoring method that uses sensors to collect data to identify the occasional state of traffic. Compared with traditional manual monitoring, intelligent monitoring systems have many advantages, such as being able to achieve continuous monitoring throughout the day, low cost and the ability to protect personal information. It is widely used in comprehensive traffic management systems to effectively reduce traffic delays and avoid secondary traffic accidents.

[0003] Traditional traffic detection methods mainly achieve event discrimination by extracting foreground targets from sequence images frame by frame from video images. Common methods include frame difference method, background difference method, optical flow method and edge detection. However, in recent years, with the improvement of artificial intelligence algorithms and the computing power of hardware equipment, traffic event detection based on machine learning has become a research hotspot. Researchers have proposed various methods based on deep learning and convolutional neural networks to achieve the detection and tracking of traffic congestion and abnormal events. Traffic event detection methods based on machine learning and deep learning have achieved significant improvements in accuracy and efficiency, providing more powerful technical support for traffic management and safety. However, the sporadic nature of abnormal events leads to the problem of uneven data distribution, making it difficult for event detection data sets to meet the independent and identically distributed assumptions of machine learning, which is prone to overfitting and insensitivity to sparse samples, resulting in a high false detection rate of abnormal events in traffic flow.

[0004] Machine learning has greatly improved the ability to detect traffic incidents and has become the mainstream research direction of current event detection. However, existing detection methods are affected by environmental factors such as lighting conditions and visibility, resulting in a high false detection rate for abnormal traffic flow events under bad weather conditions. It is urgent to develop all-weather abnormal traffic event detection methods to address the real problem of frequent traffic incidents under bad weather conditions. Summary of the invention

[0005] The present invention provides a heavy-load traffic abnormal event detection and state discrimination system to solve the defects in the prior art.

[0006] The present invention is achieved through the following technical solutions:

[0007] A heavy-load traffic abnormal event detection and state discrimination system, comprising a data acquisition and annotation module, an image processing module, a traffic state analysis module, a state judgment and abnormal event recognition module, and a transfer learning module;

[0008] The data collection and annotation module includes a data collection unit and a data annotation unit;

[0009] The data acquisition unit can collect data sets in the field using a heavy-load and high-traffic road section video detector under adverse weather conditions, and divide the data sets into scenes;

[0010] The data annotation unit can annotate the collected data set, annotate traffic accidents, spilled objects, and stranded vehicle events, so as to build a heavy-duty traffic image data set;

[0011] The image processing module includes an image enhancement unit and an image compression and reconstruction unit;

[0012] The image enhancement unit can perform image enhancement on the original data according to the traffic characteristic parameters effectively extracted from the video data under different weather conditions;

[0013] The image compression and reconstruction unit can compress and reconstruct the enhanced image to improve the image quality;

[0014] The traffic status analysis module includes a traffic bottleneck identification unit and a traffic wave analysis unit;

[0015] The traffic bottleneck identification unit can identify the road section occupied by the accident when a traffic incident occurs on the road, and determine the traffic bottleneck section;

[0016] The traffic wave analysis unit can analyze the inflow rate and traffic capacity of traffic entering the bottleneck section according to the traffic wave theory, and determine the propagation of traffic waves;

[0017] The state judgment and abnormal event recognition module includes a vehicle statistics unit, a congestion state recognition unit, a state transition analysis unit and a compressed sensing model establishment unit;

[0018] The vehicle counting unit can divide the detection area into lanes, count the number of vehicles in each lane and the number of vehicles leaving each area;

[0019] The congestion state identification unit can use the macro basic diagram of the traffic flow model to perform curve fitting, solve the curve change trend, and identify the congestion state;

[0020] The state transition analysis unit can propose a traffic state judgment method based on a Markov chain, establish a state transition matrix, and realize traffic state judgment and abnormal event recognition;

[0021] The compressed sensing model building unit can pre-process the collected road traffic video data, including denoising and reducing the number of samples to reduce the data dimension and complexity;

[0022] The transfer learning module includes a transfer learning unit:

[0023] The transfer learning unit can transfer the model trained for abnormal event detection under normal weather conditions to that under adverse weather conditions through transfer learning based on actual traffic video data and abnormal traffic flow characteristic parameter performance, so as to accelerate the deployment of heavy-load flow abnormal event detection.

[0024] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the data acquisition unit can use the multi-sensor fusion technology combining cameras and laser radars to collect data in heavy-load and high-flow sections; by setting the parameters of acquisition frequency, angle and resolution under adverse weather conditions, it is ensured that comprehensive data is obtained in various complex environments; the parameters are adjusted in real time to cope with dynamically changing weather conditions to ensure the comprehensiveness and accuracy of the data. Suppose the training sample set of abnormal events of the acquisition sensor is , then the expression of the sensor abnormal event detection model is:

[0025]

[0026] in, Represents the input data weight; Indicates the offset; Indicates the number of samples; ReLU activation function. represents the output layer weight; Represents the number of hidden layer units of the neural network; represents the target output, is the input data vector, which contains the data of the current time step and multiple historical time steps. For the time steps of sensor data, For delay The time step data, For longer historical time step data, Represents the prediction target, that is, the data of the next time step. This dataset is used to train the neural network so that it can predict the next time step based on the data of multiple time steps in the past. Predicting future sensor states , to detect abnormal events;

[0027] The model learns the historical patterns of sensors through neural networks. When the sensor data deviates, the model will not be able to accurately predict , thereby triggering anomaly detection, in order to improve computational efficiency, it can be converted into a matrix representation:

[0028]

[0029] in, represents the output vector of abnormal event detection in sensor networks; represents the matrix of neurons, Represents the weight matrix and its expression is:

[0030]

[0031] The final sensor abnormal event detection model is:

[0032]

[0033] The above formula is used to calculate the final sensor prediction value , and is used to determine whether an abnormal event has occurred.

[0034] In the above-mentioned heavy-load traffic abnormal event detection and state determination system, the image enhancement unit adopts adaptive histogram equalization technology to improve the contrast of the image, especially to significantly enhance the image details under poor light conditions.

[0035] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the image enhancement unit uses a dark channel algorithm to process video data in bad weather to improve the clarity of the image. The dark channel mathematical model expression is as follows:

[0036]

[0037] Among them, the dark channel image is the original image A property of Local window at the location Minimum channel value within , , The minimum value of

[0038] For a given pixel First, for each channel, calculate its Minimum value within ; Then, take the global minimum value of the minimum values ​​of these three channels, and you will get the dark channel image ;

[0039] At the same time, according to the atmospheric physical scattering model, by combining the color and intensity of the atmospheric light in bad weather and the scene observed on a sunny day, the color and intensity of a pixel in the image are approximately obtained, and its expression is:

[0040]

[0041] in, Indicates the deblurred image in pixels The brightness at is the original image in pixels The brightness at The transmittance is expressed in pixels The degree of attenuation of light when passing through the atmosphere; It is the atmospheric light intensity, which indicates the maximum brightness value of the image without the influence of adverse weather conditions. The higher the transmittance, the less the light attenuation and the clearer the image.

[0042] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the mathematical model of the dark channel is derived by minimum value filtering, and the estimation formula of transmittance distribution is derived. The three channels are respectively for the pixels Find the minimum value filter, the expression is:

[0043]

[0044] in, Indicates that the input image is in RGB channels Upper pixel The strength value of Indicates that in an unobstructed scene, the image is in the RGB channel Upper pixel The strength value of Indicates the atmospheric light value in the RGB channel Strength on Indicates location The transmittance at ;

[0045] Divide the atmospheric light intensity value on both sides have to:

[0046]

[0047] For RGB After the three channels are filtered for minimum value, regional minimum value filtering is performed to obtain the transmittance, which is expressed as:

[0048]

[0049] At the same time, the image enhancement parameters are adjusted dynamically in real time according to different weather conditions to ensure the best image quality under various ambient light and visibility conditions. First, the image segmentation is represented as a linear function through the local linear model of guided filtering, and the coefficients of the linear function are solved by the least squares method to minimize the filtering error. Secondly, the regularization parameter is introduced to adjust the filter effect to prevent the loss of details caused by over-filtering. Finally, the guided filtering is applied to the deblurred image to reduce noise and improve the clarity of the image, so as to obtain a more realistic and clear final blurred image. The relationship satisfied by the function is expressed as:

[0050]

[0051] in, represents the output image; represents the image pixels before filtering; and Represents the index of the pixel; and Indicates that when the center of the window is The coefficient of the function when ;

[0052] The coefficients of the linear function are solved using the least squares method, and the expression is as follows:

[0053]

[0054] in Indicates in the local window Solving Linear Functions Internally The coefficient of and The loss function includes the sum of squares of the residuals and a regularization term;

[0055] The residual sum of squares is used to measure the difference between the linear function prediction value and the actual observation value, that is, ,in is the pixel value in the local window, is the corresponding true observation value. The goal of this part of the loss function is to minimize the fitting error of the linear function to the observed data.

[0056] The regularization term is used to control the coefficient The size of to prevent overfitting; the regularization term is ,in is the regularization parameter. The goal of this part of the loss function is to keep the coefficients Smoothness;

[0057] Through the least squares method, the coefficients can be obtained. and The loss function is expressed as:

[0058]

[0059] in, express The average value of express The variance of Indicates the number of pixels; Represents the image to be filtered The mean of

[0060] The above analytical solution ensures that the output is within the local window. With guidance Keep the same structural changes while keeping it as close to the original image as possible The average brightness and contrast, Numerical expression and Degree of relevance: If and Increase and decrease in the window. Close to 1, the output almost replicates the guiding image structure; if the correlation between the two is weak or is almost constant, then Tends to 0, the output is approximately smoothed by taking the window mean. Then adjust the local brightness so that The mean value within the window is equal to The mean of .

[0061] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the image compression and reconstruction unit uses a variational autoencoder (VAE) to compress the video data. During the compression process, the key frames of the video are retained and the redundant frames are deleted, thereby significantly reducing the amount of data; assuming that the real-valued signal Length is , which can be expressed as , ,but Available base describe:

[0062]

[0063] in, Indicates a Wiki matrix, basis vectors are ; represents the sparse transformation coefficient matrix, which represents the sparse coefficient representation of the signal under the basis vector;

[0064] The signal is projected onto the perception matrix after sparse transformation ( for ), generating the observation vector , its mathematical expression is:

[0065]

[0066] in, is the observation vector, with dimension , by the perception matrix Acting on a signal get;

[0067] Comprehensively obtained:

[0068]

[0069] in, represents the composite matrix; for dimensional measurement matrix; Indicates basis vectors; is a sparse coefficient vector;

[0070] Simultaneous sensing matrix Satisfy the RIP (RestrictedIsometryProperty) condition:

[0071]

[0072] According to the observation vector expression, we can know is sparse, and The conditions can be met by the minimum expression of the observation vector Norm is used to solve, according to Reconstruct the original signal , the expression is:

[0073]

[0074] in, The norm solution process is an NP-hard problem. Under certain conditions, Norm and The norm is equivalent and its expression is:

[0075]

[0076] The operation of the image compression and reconstruction unit includes the following steps:

[0077] Step 1): Input: Dimensional matrix , and the sparsity of the signal ;

[0078] Step 2): Output: Index set , the reconstructed signal ,margin ;

[0079] Step 3): , , number of iterations ;

[0080] Step 4): Search for the new index , so that it satisfies the condition ;

[0081] Step 5): Set , using the least squares method to find the approximate solution of the signal ;

[0082] Step 6): ;

[0083] Step 7): The number of iterations increases, if the conditions are met , then go to step 4) and continue iterating; otherwise, end.

[0084] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the traffic bottleneck identification unit extracts key characteristic parameters of vehicle flow, vehicle speed, and vehicle density based on the collected video data, analyzes the capacity of the vehicle flow when entering the bottleneck section, evaluates the impact of the bottleneck on the traffic flow, and determines the propagation of the traffic wave;

[0085] The traffic bottleneck identification unit uses the K-means clustering algorithm to analyze the distribution and characteristics of traffic flow and identify traffic bottleneck sections; by combining historical data with real-time monitoring data, the threshold of bottleneck identification is dynamically adjusted to improve the accuracy and sensitivity of identification.

[0086] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the traffic wave analysis unit can analyze the impact of traffic events on traffic based on the traffic wave propagation model of the space-time graph;

[0087] Combined with the fluid mechanics model, the propagation process of traffic waves on the road is simulated, and its propagation speed and impact range in different sections are predicted. Based on the Lighthill-Whitham-Richards (LWR) model, the continuity equation and momentum equation of traffic flow are established to describe the density change and speed change of traffic flow: ,in, represents the vehicle density, Indicates vehicle speed.

[0088] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the state transition analysis unit can construct a Markov chain model of the traffic state. The Markov chain model predicts the change trend of the traffic state by updating the transfer matrix in real time, and combines the Bayesian reasoning method to identify abnormal events, and divides the traffic state into several discrete states. These states may include but are not limited to: normal flow, slow flow, congestion, accident occurrence, vehicle spillage, stranded vehicles, and these discrete states are used as the state set of the Markov chain model. , define the state transfer matrix ,in Indicates that the system is in state Transfer to state The probability of satisfying the following conditions:

[0089] Setting the prior probability of different abnormal events ,in Indicates The occurrence of abnormal events;

[0090] Collect real-time traffic data, which are used as observations ; Establish a likelihood function based on the impact of different abnormal events on traffic conditions , which indicates that in the event of an abnormal When it happens, the data is observed probability;

[0091] Using Bayes' theorem, calculate the posterior probability , that is, when the data is observed After the abnormal event Probability of occurrence:

[0092]

[0093] in, Indicates that when the data is observed After the event The posterior probability of occurrence; Indicates a hypothetical event Under the premise of occurrence, observation data The likelihood of occurrence is called the likelihood function; Indicates an event The prior probability of occurrence indicates that without any new data When the event occurs, the subjective belief that the event occurred; Represents the combined impact of all possible events, ensuring the normalization of all posterior probabilities so that the sum of the probabilities of all possible events is 1;

[0094] An event with a higher posterior probability is considered to be more likely to occur;

[0095] By calculating the posterior probability of each abnormal event in real time, the most likely abnormal event is identified.

[0096] In the heavy-load traffic abnormal event detection and state discrimination system as described above, the transfer learning unit applies the abnormal event detection model trained under normal weather conditions to adverse weather conditions through the transfer learning method, and fine-tunes and optimizes the model under the new weather conditions to ensure that the model can efficiently and accurately detect abnormal events under various weather conditions; through transfer learning, the deployment process of the abnormal event detection model is accelerated, the time and resource consumption of retraining the model are reduced, and the flexibility and adaptability of the system are improved.

[0097] The advantages of the present invention are that the present invention can reduce the false detection rate of abnormal traffic flow events under bad weather conditions, thereby being able to detect abnormal traffic events around the clock, thereby being able to reduce the probability of traffic events occurring under bad weather conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0099] Figure 1 It is a structural block diagram of the present invention;

[0100] Figure 2 is a flow chart of the present invention;

[0101] Figure 3 is a flow chart of the traffic status analysis module of the present invention;

[0102] Figure 4 It is a flow chart of the state judgment and abnormal event identification module of the present invention. DETAILED DESCRIPTION

[0103] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0104] like Figure 1 and Figure 2 As shown, a heavy-load traffic abnormal event detection and state discrimination system includes a data acquisition and annotation module, an image processing module, a traffic state analysis module, a state judgment and abnormal event recognition module, and a transfer learning module;

[0105] The data collection and annotation module includes a data collection unit and a data annotation unit;

[0106] The data acquisition unit can collect data sets in the field using a heavy-load and high-traffic road section video detector under adverse weather conditions, and divide the data sets into scenes;

[0107] The data annotation unit can annotate the collected data sets, annotate traffic accidents, spilled objects, and stranded vehicle events to build a heavy-duty traffic image data set; through annotation, a detailed and comprehensive heavy-duty traffic image data set is constructed to provide basic data support for subsequent image processing and analysis.

[0108] The image processing module includes an image enhancement unit and an image compression and reconstruction unit; the image distortion caused by haze or precipitation is reduced by an adaptive image enhancement method, image noise is removed, image contrast and brightness are adjusted, and image quality is improved.

[0109] The image enhancement unit can enhance the original data according to the traffic characteristic parameters effectively extracted from the video data under different weather conditions; it uses advanced image enhancement technology to process the video data to improve the clarity and quality of the image, ensuring that high-quality images can be obtained under different weather conditions;

[0110] The image compression and reconstruction unit can compress and reconstruct the enhanced image to improve the image quality, and use the compression and reconstruction algorithm to compress the processed video data to reduce the pressure of data storage and transmission. Under the premise of ensuring the image quality, the data volume is reduced to the maximum extent.

[0111] like Figure 3 As shown, the traffic status analysis module includes a traffic bottleneck identification unit and a traffic flow wave analysis unit;

[0112] The traffic bottleneck identification unit can identify the road section occupied by the accident when a traffic incident occurs on the road, and determine the traffic bottleneck section; use the real-time video stream captured by the camera to conduct a comprehensive analysis of key parameters such as vehicle density, speed and flow direction on the road, and use the image enhancement method of multi-scale structural similarity loss to analyze the real-time video captured by the road monitoring camera, detect areas with abnormally high traffic flow density, and quickly and accurately locate the road section affected by the accident

[0113] The traffic wave analysis unit can analyze the inflow rate and traffic capacity of traffic entering the bottleneck section according to the traffic wave theory, and determine the propagation of traffic waves; based on the traffic theory, a traffic flow model based on macroscopic theory is established to describe the movement behavior of heavy-loaded vehicles on the road, and historical data of heavy-loaded and high-flow section video detection is used, such as the density-speed-flow relationship, to analyze the inflow rate and traffic capacity of heavy-loaded traffic entering the bottleneck section, and simulate the propagation process of traffic flow on the road in combination with real-time data and traffic flow models to predict the propagation of traffic waves.

[0114] like Figure 4 As shown, the state judgment and abnormal event recognition module includes a vehicle statistics unit, a congestion state recognition unit, a state transition analysis unit and a compressed sensing model establishment unit;

[0115] The vehicle counting unit can divide the detection area into lanes, count the number of vehicles in each lane and the number of vehicles leaving each community; use image enhancement methods to process the video captured by the road monitoring camera, use deep learning target detection algorithm (YOLO) to detect and count vehicles, detect and track vehicles, and thus count the number of vehicles in each lane and the number of vehicles leaving each community; use convolutional neural networks to detect and classify vehicles to identify and count the number of vehicles in traffic.

[0116] The congestion state identification unit can use the macro basic diagram of the traffic flow model to perform curve fitting, solve the curve change trend, and identify the congestion state; use a macro traffic flow model, such as a flow density-speed relationship model, to analyze the real-time monitored heavy-load vehicle density and speed data to identify the congestion state; based on the historical data of video detection of heavy-loaded and high-flow sections, use a support vector machine to train a classification model of the congestion state, combine the road network topology and traffic flow dynamic characteristics, design congestion indicators and thresholds, and judge the traffic congestion state through real-time monitoring data.

[0117] The state transition analysis unit can propose a traffic state judgment method based on Markov chain, establish a state transition matrix, realize traffic state judgment and abnormal event recognition; establish a Markov chain model to describe the transition relationship between different traffic states, including unobstructed, light congestion, moderate congestion and severe congestion, etc., based on real-time monitoring data, use statistical methods and time series analysis technology to update the state transition probability of the Markov chain, and realize dynamic recognition and analysis of traffic states.

[0118] The compressed sensing model building unit can pre-process the collected road traffic video data, including denoising and reducing the number of samples to reduce the data dimension and complexity; then, using compressed sensing theory and methods, design and train image encoding and decoding models, extract key features from video data, and convert them into efficient sparse representations, establish an abnormal event image semantic representation model based on compressed sensing, and better capture the image semantic information when abnormal events occur.

[0119] The transfer learning module includes a transfer learning unit:

[0120] The transfer learning unit can transfer the model trained for abnormal event detection under normal weather conditions to that under adverse weather conditions through transfer learning based on actual traffic video data and abnormal traffic flow characteristic parameter performance, so as to accelerate the deployment of heavy-load flow abnormal event detection.

[0121] Specifically, the data acquisition unit described in this embodiment can use a multi-sensor fusion technology combining cameras and lidar to collect data in heavy-load and high-flow sections; thereby, it can collect highly robust data under different weather conditions (such as rainy days, snowy days, foggy days, etc.); by setting the parameters of acquisition frequency, angle and resolution under adverse weather conditions, it is ensured that comprehensive data is obtained in various complex environments; the parameters are adjusted in real time to cope with dynamically changing weather conditions to ensure the comprehensiveness and accuracy of the data. Suppose the training sample set of abnormal events of the acquisition sensor is , then the expression of the sensor abnormal event detection model is:

[0122]

[0123] in, Represents the input data weight; Indicates the offset; Indicates the number of samples; ReLU activation function. represents the output layer weight; Represents the number of hidden layer units of the neural network; represents the target output, is the input data vector, which contains the data of the current time step and multiple historical time steps. For the time steps of sensor data, For delay The time step data, For longer historical time step data, Represents the prediction target, that is, the data of the next time step. This dataset is used to train the neural network so that it can predict the next time step based on the data of multiple time steps in the past. Predicting future sensor states , to detect abnormal events;

[0124] The model learns the historical patterns of sensors through neural networks. When the sensor data deviates, the model will not be able to accurately predict , thereby triggering anomaly detection, in order to improve computational efficiency, it can be converted into a matrix representation:

[0125]

[0126] in, represents the output vector of abnormal event detection in sensor networks; represents the matrix of neurons, Represents the weight matrix and its expression is:

[0127]

[0128] The final sensor abnormal event detection model is:

[0129]

[0130] The above formula is used to calculate the final sensor prediction value , and is used to determine whether an abnormal event has occurred.

[0131] During the data collection process, the status of all sensors is monitored in real time, including power consumption, temperature, signal strength, etc., to ensure that the quality of the collected data meets the requirements. When an abnormality is found, an alarm is issued in time or the collection strategy is adjusted;

[0132] Develop a set of semi-automatic annotation tools, which combine automatic detection algorithms to perform preliminary annotation of video data, and then manually perform precise correction of key frames and events;

[0133] Establish detailed labeling standards to cover various types of traffic events, such as vehicle driving conditions, traffic accidents, spilled objects, stranded vehicles, etc.;

[0134] Use crowdsourcing platforms to mobilize multiple labelers to participate in data labeling, and ensure the high quality and accuracy of labeled data through cross-validation and consistency checks.

[0135] Specifically, the image enhancement unit described in this embodiment adopts adaptive histogram equalization (Contrast Limited Adaptive Histogram Equalization, CLAHE) technology to improve the contrast of the image, especially significantly enhance the image details under poor lighting conditions; the adaptive histogram equalization technology can better highlight the key visual information and improve the detection accuracy of the algorithm.

[0136] More specifically, the image enhancement unit described in this embodiment uses a dark channel algorithm to process video data in bad weather to improve the clarity of the image. In bad weather, the dark channel algorithm improves the video quality by restoring the true color and details of the image, thereby improving the accuracy of traffic event detection. The dark channel mathematical model expression is as follows:

[0137]

[0138] Among them, the dark channel image is the original image A property of Local window at the location Minimum channel value within , , The minimum value of

[0139] For a given pixel First, for each channel, calculate its Minimum value within ; Then, take the global minimum value of the minimum values ​​of these three channels, and you will get the dark channel image ;

[0140] At the same time, according to the atmospheric physical scattering model, by combining the color and intensity of the atmospheric light in bad weather and the scene observed on a sunny day, the color and intensity of a pixel in the image are approximately obtained, and its expression is:

[0141]

[0142] in, Indicates the deblurred image in pixels The brightness at is the original image in pixels The brightness at The transmittance is expressed in pixels The degree of attenuation of light when passing through the atmosphere; It is the atmospheric light intensity, which indicates the maximum brightness value of the image without the influence of adverse weather conditions. The higher the transmittance, the less the light attenuation and the clearer the image.

[0143] Brightness of the deblurred image is the original image and atmospheric light intensity The weighted average of If the transmittance is close to 1, it means that the light is almost not attenuated by the fog. Mainly foggy images If the transmittance is close to 0, it means that the light is greatly attenuated. Close to atmospheric light intensity .

[0144] More specifically, the mathematical model of the dark channel described in this embodiment is derived by minimum filtering, and the estimation formula of the transmittance distribution is derived. The three channels are respectively for the pixels Find the minimum value filter, the expression is:

[0145]

[0146] in, Indicates that the input image is in RGB channels Upper pixel The strength value of Indicates that in an unobstructed scene, the image is in the RGB channel Upper pixel The strength value of Indicates the atmospheric light value in the RGB channel Strength on Indicates location The transmittance at ;

[0147] Divide the atmospheric light intensity value on both sides have to:

[0148]

[0149] For RGB After the three channels are filtered for minimum value, regional minimum value filtering is performed to obtain the transmittance, which is expressed as:

[0150]

[0151] At the same time, the image enhancement parameters are adjusted dynamically in real time according to different weather conditions to ensure the best image quality under various ambient light and visibility conditions. First, the image segmentation is represented as a linear function through the local linear model of guided filtering, and the coefficients of the linear function are solved by the least squares method to minimize the filtering error. Secondly, the regularization parameter is introduced to adjust the filter effect to prevent the loss of details caused by over-filtering. Finally, the guided filtering is applied to the deblurred image to reduce noise and improve the clarity of the image, so as to obtain a more realistic and clear final blurred image. The relationship satisfied by the function is expressed as:

[0152]

[0153] in, represents the output image; represents the image pixels before filtering; and Represents the index of the pixel; and Indicates that when the center of the window is The coefficient of the function when ;

[0154] The coefficients of the linear function are solved using the least squares method, and the expression is as follows:

[0155]

[0156] in Indicates in the local window Solving Linear Functions Internally The coefficient of and The loss function includes the sum of squares of the residuals and a regularization term;

[0157] The residual sum of squares is used to measure the difference between the linear function prediction value and the actual observation value, that is, ,in is the pixel value in the local window, is the corresponding true observation value. The goal of this part of the loss function is to minimize the fitting error of the linear function to the observed data.

[0158] The regularization term is used to control the coefficient The size of to prevent overfitting; the regularization term is ,in is the regularization parameter. The goal of this part of the loss function is to keep the coefficients Smoothness;

[0159] Through the least squares method, the coefficients can be obtained. and The loss function is expressed as:

[0160]

[0161] in, express The average value of express The variance of Indicates the number of pixels; Represents the image to be filtered The mean of

[0162] The above analytical solution ensures that the output is within the local window. With guidance Keep the same structural changes while keeping it as close to the original image as possible The average brightness and contrast, Numerical expression and Degree of relevance: If and Increase and decrease in the window. Close to 1, the output almost replicates the guiding image structure; if the correlation between the two is weak or is almost constant, then Tends to 0, the output is approximately smoothed by taking the window mean. Then adjust the local brightness so that The mean value within the window is equal to The mean of .

[0163] In the above-mentioned heavy-load traffic abnormal event detection and state discrimination system, the image compression and reconstruction unit uses a variational autoencoder (VAE) to compress the video data. During the compression process, the key frames of the video are retained and the redundant frames are deleted, thereby significantly reducing the amount of data; assuming that the real-valued signal Length is , which can be expressed as , ,but Available base describe:

[0164]

[0165] in, Indicates a Wiki matrix, basis vectors are ; represents the sparse transformation coefficient matrix, which represents the sparse coefficient representation of the signal under the basis vector;

[0166] The signal is projected onto the perception matrix after sparse transformation ( for ), generating the observation vector , its mathematical expression is:

[0167]

[0168] in, is the observation vector, with dimension , by the perception matrix Acting on a signal get;

[0169] Comprehensively obtained:

[0170]

[0171] in, represents the composite matrix; for dimensional measurement matrix; Indicates basis vectors; is a sparse coefficient vector;

[0172] Simultaneous sensing matrix Satisfy the RIP (Restricted Isometry Property) condition:

[0173]

[0174] According to the observation vector expression, we can know is sparse, and The conditions can be met by the minimum expression of the observation vector Norm is used to solve, according to Reconstruct the original signal , the expression is:

[0175]

[0176] in, The norm solution process is an NP-hard problem. Under certain conditions, Norm and The norm is equivalent and its expression is:

[0177]

[0178] The operation of the image compression and reconstruction unit includes the following steps:

[0179] Step 1): Input: Dimensional matrix , and the sparsity of the signal ;

[0180] Step 2): Output: Index set , the reconstructed signal ,margin ;

[0181] Step 3): , , number of iterations ;

[0182] Step 4): Search for the new index , so that it satisfies the condition ;

[0183] Step 5): Set , using the least squares method to find the approximate solution of the signal ;

[0184] Step 6): ;

[0185] Step 7): The number of iterations increases, if the conditions are met , then go to step 4) and continue iterating; otherwise, end.

[0186] Furthermore, the traffic bottleneck identification unit described in this embodiment extracts key characteristic parameters of vehicle flow, vehicle speed, and vehicle density based on the collected video data, analyzes the capacity of the vehicle flow when entering the bottleneck section, evaluates the impact of the bottleneck on the traffic flow, and determines the propagation of the traffic wave;

[0187] The traffic bottleneck identification unit uses the K-means clustering algorithm to analyze the distribution and characteristics of traffic flow and identify traffic bottleneck sections; by combining historical data and real-time monitoring data, the threshold of bottleneck identification is dynamically adjusted to improve the accuracy and sensitivity of identification. The system continuously monitors traffic flow data and dynamically adjusts the parameters and models of bottleneck identification through a real-time feedback mechanism to ensure the accuracy of bottleneck identification results under different traffic conditions.

[0188] Furthermore, the traffic wave analysis unit described in this embodiment can analyze the impact of traffic events on traffic flow based on the traffic wave propagation model of the space-time graph;

[0189] Combined with the fluid mechanics model, the propagation process of traffic waves on the road is simulated, and its propagation speed and impact range in different sections are predicted. Based on the Lighthill-Whitham-Richards (LWR) model, the continuity equation and momentum equation of traffic flow are established to describe the density change and speed change of traffic flow: ,in, represents the vehicle density, Indicates vehicle speed.

[0190] The space-time graph model is integrated with the fluid dynamics model to form a joint analysis framework. The space-time graph model is used to capture the dynamic changes of traffic waves, and the fluid dynamics model is used to simulate the physical propagation process of waves.

[0191] The traffic wave propagation model based on the space-time graph is used to analyze the impact of traffic events on traffic flow. The model can accurately describe the propagation process of traffic waves in different sections of the road, helping to understand the impact of traffic events on the entire traffic system. Combined with the fluid mechanics model, the propagation process of traffic waves on the road is simulated to predict its propagation speed and impact range in different sections of the road. Through regression analysis of historical data, the model parameters are calibrated to improve the accuracy and reliability of the model.

[0192] Furthermore, the state transition analysis unit described in this embodiment can construct a Markov chain model of the traffic state. The Markov chain model predicts the changing trend of the traffic state by updating the transfer matrix in real time, and combines the Bayesian reasoning method to identify abnormal events, and divides the traffic state into several discrete states. These states may include but are not limited to: normal flow, slow flow, congestion, accidents, vehicle spillage, and stranded vehicles. These discrete states are used as the state set of the Markov chain model. , define the state transfer matrix ,in Indicates that the system is in state Transfer to state The probability of satisfying the following conditions:

[0193] Using historical traffic data and real-time traffic data, the transition frequencies between states are counted to estimate the state transition matrix. The state transition probability is calculated by using the frequency estimation method or the maximum likelihood estimation method to ensure the accuracy of the transfer matrix;

[0194] As traffic conditions change, new traffic data is collected in real time, and the state transfer matrix is ​​adjusted and updated accordingly ;

[0195] Through continuous data input, the state transfer matrix is ​​dynamically optimized to always reflect the current traffic state change trend;

[0196] Using the Markov chain model, according to the current traffic status and the transfer matrix , predict the traffic status of the next time step ;

[0197] Through multi-step prediction, a traffic status sequence for multiple time steps in the future can be generated to predict the changing trend of the traffic status.

[0198] Use Markov chain models and Bayesian reasoning to quickly identify abnormal events, such as traffic accidents, vehicle spillage, stranded vehicles, etc.

[0199] Set the prior probability of different abnormal events (such as traffic accidents, spilled objects, stranded vehicles, etc.) ,in Indicates The occurrence of abnormal events;

[0200] The prior probability of an abnormal event reflects the initial estimate of the occurrence of the event when there is no observed data;

[0201] Collect real-time traffic data, such as sudden drop in vehicle speed, abnormal increase in vehicle density, etc., and use these data as observations ; Establish a likelihood function based on the impact of different abnormal events on traffic conditions , which indicates that in the event of an abnormal When it happens, the data is observed probability;

[0202] Using Bayes' theorem, calculate the posterior probability , that is, when the data is observed After the abnormal event Probability of occurrence:

[0203]

[0204] in, Indicates that when the data is observed After the event The posterior probability of occurrence; Indicates a hypothetical event Under the premise of occurrence, observation data The likelihood of occurrence is called the likelihood function; Indicates an event The prior probability of occurrence indicates that without any new data When the event occurs, the subjective belief that the event occurred; Represents the combined impact of all possible events, ensuring the normalization of all posterior probabilities so that the sum of the probabilities of all possible events is 1;

[0205] An event with a higher posterior probability is considered to be more likely to occur;

[0206] By calculating the posterior probability of each abnormal event in real time, the most likely abnormal event can be identified. When the posterior probability of an event exceeds the preset threshold, the system generates an alarm signal to prompt traffic management personnel to take corresponding measures; through real-time monitoring and dynamic adjustment, the identification process can be continuously optimized to ensure timely and accurate detection of traffic abnormal events.

[0207] Furthermore, the transfer learning unit described in this embodiment applies the abnormal event detection model trained under normal weather conditions to adverse weather conditions through the transfer learning method, and fine-tunes and optimizes the model under the new weather conditions to ensure that the model can efficiently and accurately detect abnormal events under various weather conditions; through transfer learning, the deployment process of the abnormal event detection model is accelerated, the time and resource consumption of retraining the model is reduced, and the flexibility and adaptability of the system are improved.

[0208] Use a pre-trained model (anomaly detection model trained under normal weather conditions) and fine-tune it under adverse weather conditions through transfer learning;

[0209] Collect new data under adverse weather conditions and annotate them as a fine-tuning dataset;

[0210] Adopt a fine-tuning strategy to adjust only some parameters of the model, such as the output layer weight, to maintain the feature extraction capability of the pre-trained model;

[0211] Through distillation learning technology, the knowledge of complex models is transferred to lightweight models to accelerate the reasoning speed of the models;

[0212] Use edge computing technology to deploy some models on edge devices to reduce data transmission delays and improve the system's real-time detection capabilities;

[0213] Use the distributed training framework to speed up the model training process and use the distributed deployment strategy to achieve rapid launch.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A heavy-load traffic abnormal event detection and state identification system, characterized by: It includes data collection and annotation module, image processing module, traffic status analysis module, status judgment and abnormal event recognition module and transfer learning module; The data collection and annotation module includes a data collection unit and a data annotation unit; The data acquisition unit can collect data sets in the field using a heavy-load and high-traffic road section video detector under adverse weather conditions, and divide the data sets into scenes; The data annotation unit can annotate the collected data set, annotate traffic accidents, spilled objects, and stranded vehicle events, so as to build a heavy-duty traffic image data set; The image processing module includes an image enhancement unit and an image compression and reconstruction unit; The image enhancement unit can perform image enhancement on the original data according to the traffic characteristic parameters effectively extracted from the video data under different weather conditions; The image compression and reconstruction unit can compress and reconstruct the enhanced image to improve the image quality; The traffic status analysis module includes a traffic bottleneck identification unit and a traffic wave analysis unit; The traffic bottleneck identification unit can identify the road section occupied by the accident when a traffic incident occurs on the road, and determine the traffic bottleneck section; The traffic wave analysis unit can analyze the inflow rate and traffic capacity of traffic entering the bottleneck section according to the traffic wave theory, and determine the propagation of traffic waves; The state judgment and abnormal event recognition module includes a vehicle statistics unit, a congestion state recognition unit, a state transition analysis unit and a compressed sensing model establishment unit; The vehicle counting unit can divide the detection area into lanes, count the number of vehicles in each lane and the number of vehicles leaving each area; The congestion state identification unit can use the macro basic diagram of the traffic flow model to perform curve fitting, solve the curve change trend, and identify the congestion state; The state transition analysis unit can propose a traffic state judgment method based on a Markov chain, establish a state transition matrix, and realize traffic state judgment and abnormal event recognition; The compressed sensing model building unit can pre-process the collected road traffic video data, including denoising and reducing the number of samples to reduce the data dimension and complexity; The transfer learning module includes a transfer learning unit: The transfer learning unit can transfer the model trained for abnormal event detection under normal weather conditions to that under adverse weather conditions through transfer learning based on actual traffic video data and abnormal traffic flow characteristic parameter performance, so as to accelerate the deployment of heavy-load flow abnormal event detection.

2. A heavy-load traffic abnormal event detection and state identification system according to claim 1, characterized in that: The data acquisition unit can use the multi-sensor fusion technology combining cameras and laser radar to collect data in heavy-load and high-traffic sections; by setting the parameters of acquisition frequency, angle and resolution under adverse weather conditions, it ensures that comprehensive data is obtained in various complex environments; adjust the parameters in real time to cope with dynamically changing weather conditions to ensure the comprehensiveness and accuracy of the data. The training sample set of abnormal events of the acquisition sensor is Then the expression of the sensor abnormal event detection model is: Among them, a i represents the input data weight; b i represents the bias; k represents the number of samples; f(·) represents the ReLU activation function; β i represents the output layer weight; L represents the number of hidden layer units of the neural network; t represents the target output, X i is the input data vector, which contains the data of the current time step and multiple historical time steps, x i is the sensor data at the i-th time step, x i-τ is the data delayed by τ time steps, x i-(m-1)τ For longer historical time step data, t i =x i+1 Represents the prediction target, that is, the data of the next time step. This dataset is used to train the neural network so that it can predict the next time step based on the data X of the past multiple time steps. i Predict future sensor states t i , to detect abnormal events; The model learns the historical patterns of sensors through neural networks. When the sensor data deviates, the model will not be able to accurately predict t i , thereby triggering anomaly detection, in order to improve computational efficiency, it can be converted into a matrix representation: H k b k =T k Among them, T k H represents the output vector of abnormal event detection in sensor networks; k represents the matrix of neurons, β k Represents the weight matrix and its expression is: The final sensor abnormal event detection model is: The above formula is used to calculate the final sensor prediction value t and to determine whether an abnormal event occurs.

3. The heavy-load traffic abnormal event detection and state identification system according to claim 1 is characterized by: The image enhancement unit uses an adaptive histogram equalization technique to improve the contrast of the image and significantly enhance the image details under poor light conditions.

4. A heavy-load traffic abnormal event detection and state identification system according to claim 3, characterized in that: The image enhancement unit uses a dark channel algorithm to process video data in bad weather to improve the clarity of the image. The dark channel mathematical model expression is as follows: Among them, the dark channel image J dark (x) is an attribute of the original image J, which represents the minimum value of the minimum channel value r, g, b in the local window Ω(x) at the position of pixel x, and is used to estimate the atmospheric light intensity and transmittance; J c (y) represents the brightness value of the haze-free image in the r, g, and b channels; For a given pixel x, first calculate the minimum value J of each channel within the local window Ω(x) c (x); Then, the minimum values ​​of these three channels are taken as the global minimum, and the dark channel image J is obtained. dark (x); At the same time, according to the atmospheric physical scattering model, by combining the color and intensity of the atmospheric light in bad weather and the scene observed on a sunny day, the color and intensity of a pixel in the image are approximately obtained, and its expression is: I(x)=J(x)t(x)+A(1-t(x)) Among them, I(x) represents the brightness of the deblurred image at pixel x; J(x) is the brightness of the original image at pixel x; t(x) is the transmittance, which represents the attenuation of light at pixel x when it passes through the atmosphere; A is the atmospheric light intensity, which represents the maximum brightness value of the image without the influence of adverse weather conditions; the higher the transmittance, the less the light attenuation and the clearer the image.

5. A heavy-load traffic abnormal event detection and state identification system according to claim 4, characterized in that: The mathematical model of the dark channel is subjected to minimum filtering derivation, and the estimation formula of transmittance distribution is derived. The minimum filtering is performed on the pixel y in the three channels of RGB c∈(r,g,b), and the expression is: Among them, I c (y) represents the intensity value of pixel y on RGB channel c of the input image; J c (y) represents the intensity value of pixel y on RGB channel c of the image in an unobstructed scene; A c (y) represents the intensity of the atmospheric light value on RGB channel c; t(x) represents the transmittance at position x; Divide the atmospheric light intensity value A on both sides c have to: After performing minimum filtering on the three channels (r, g, b) of RGB, regional minimum filtering is performed to obtain the transmittance, which is expressed as: At the same time, the image enhancement parameters are adjusted dynamically in real time according to different weather conditions to ensure the best image quality under various ambient light and visibility conditions. First, the image segmentation is represented as a linear function through the local linear model of guided filtering, and the coefficients of the linear function are solved by the least squares method to minimize the filtering error. Secondly, the regularization parameter is introduced to adjust the filter effect to prevent the loss of details caused by over-filtering. Finally, the guided filtering is applied to the deblurred image to reduce noise and improve the clarity of the image, so as to obtain a more realistic and clear final blurred image. The relationship satisfied by the function is expressed as: Among them, q i Represents the output image; I i represents the image pixel before filtering; i and k represent the index of the pixel; a k and b k Represents the coefficient of the function when the center of the window is located at k; The coefficients of the linear function are solved using the least squares method, and the expression is as follows: Where E(a k ,b k ) indicates that in the local window ω k Solve the linear function a k I i +b k The coefficient a k and b k The loss function includes the sum of squares of the residuals and a regularization term; The residual sum of squares is used to measure the difference between the linear function prediction value and the actual observation value, that is, (a k I i +b k -p i ) 2 , where I i is the pixel value in the local window, p i is the corresponding true observation value. The goal of this part of the loss function is to minimize the fitting error of the linear function to the observed data. The regularization term is used to control the coefficient a k The size of , to prevent overfitting; the regularization term is Where ε is the regularization parameter, the goal of this part of the loss function is to maintain the smoothness of the coefficient ε; By calculating and deriving through the least squares method, we can finally get the coefficient a k and b k The loss function is expressed as: Among them, μ k represents the average value of I; represents the variance of I; |ω| represents the number of pixels; Represents the mean value of the image p to be filtered; The above analytical solution ensures that the output q and the guide I maintain the same structural changes in the local window, while being as close to the average brightness and contrast of the original image p as possible. k The value reflects the correlation between p and I: if I and p increase and decrease together in the window, a k Close to 1, the output almost replicates the guided image structure; if the correlation between the two is weak or I is almost constant, then a k tends to 0, the output is approximately smoothed by taking the window mean, b k Then adjust the local brightness so that the mean value of q in the window is equal to the mean value of p.

6. The heavy-load traffic abnormal event detection and state identification system according to claim 1 is characterized by: The image compression and reconstruction unit uses a variational autoencoder to compress video data. During the compression process, the key frames of the video are retained and redundant frames are deleted, thereby significantly reducing the amount of data. Assuming that the length of the real-valued signal b is N, it can be expressed as b(n), n∈{1,2,…,N}, then b can be obtained using the basis ψ=(ψ1,ψ2,…,ψ N )describe: Among them, ψ represents an N×N basis matrix, and the basis vector is ψ i ; z represents the sparse transformation coefficient matrix, which represents the sparse coefficient representation of the signal under the basis vector; After sparse transformation, the signal is projected onto the perception matrix Φ to generate the observation vector y, which is mathematically expressed as: y=Φb Among them, y is the observation vector with dimension M, which is obtained by applying the perception matrix Φ to the signal b; In summary: Wherein, A=Φψ represents the synthesis matrix; Φ is the M×N (M<<N) dimensional measurement matrix; ψ i represents the i-th basis vector; z is the sparse coefficient vector; At the same time, the perception matrix Φ satisfies the Restricted Isometry Property condition: According to the observation vector expression, b is sparse and Φ satisfies the condition. The original signal b can be reconstructed according to y by solving the minimum l0 norm of the observation vector expression. The expression is: Among them, the l0 norm solution process is an NP-hard problem. Under certain conditions, the l1 norm and l0 norm are equivalent, and their expressions are: The operation of the image compression and reconstruction unit includes the following steps: Step 1): Input: M×N dimensional matrix Φ, y and signal sparsity K; Step 2): Output: Index set Λ t , the reconstructed signal Residue t ; Step 3): r0 = y, Iteration number t = 1; Step 4): Search for the new index J t , so that it satisfies the condition Step 5): Assume Λ t =Λ t-1 ∪{J t }, use the least squares method to find the approximate solution of the signal Step 6): Step 7): The number of iterations increases. If the condition t<k is met, go to step 4) and continue iterating; otherwise, end.

7. The heavy-load traffic abnormal event detection and state identification system according to claim 1 is characterized by: The traffic bottleneck identification unit extracts key characteristic parameters of vehicle flow, vehicle speed, and vehicle density based on the collected video data, analyzes the capacity of the vehicle flow when entering the bottleneck section, evaluates the impact of the bottleneck on the traffic flow, and determines the propagation of the traffic wave; The traffic bottleneck identification unit uses the K-means clustering algorithm to analyze the distribution and characteristics of traffic flow and identify traffic bottleneck sections; by combining historical data with real-time monitoring data, the bottleneck identification threshold is dynamically adjusted to improve the accuracy and sensitivity of identification.

8. The heavy-load traffic abnormal event detection and state identification system according to claim 1 is characterized by: The traffic wave analysis unit can analyze the impact of traffic events on traffic flow based on the traffic wave propagation model of the space-time graph; Combined with the fluid mechanics model, the propagation process of traffic waves on the road is simulated, and its propagation speed and impact range in different sections are predicted. Based on the Lighthill-Whitham-Richards model, the continuity equation and momentum equation of traffic flow are established to describe the density change and speed change of traffic flow: Among them, ρ represents the vehicle density and v represents the vehicle speed.

9. The heavy-load traffic abnormal event detection and state identification system according to claim 1, characterized in that: The state transition analysis unit can construct a Markov chain model of the traffic state. The Markov chain model predicts the changing trend of the traffic state by updating the transfer matrix in real time, and combines the Bayesian reasoning method to identify abnormal events, and divides the traffic state into several discrete states. These states include: normal flow, slow flow, congestion, accidents, vehicle spillage, and stranded vehicles. These discrete states are used as the state set S = {S1, S2, ..., S n }, define the state transfer matrix P, where P ij Indicates that the system changes from state S i Transfer to state S i The probability of satisfying the following conditions: P ij =Pr(S t+1 =S j |S t =S i ); Set the prior probability Pr(H i ), where H i Indicates the occurrence of the i-th abnormal event; Collect real-time traffic data, which are used as observations D; according to the impact of different abnormal events on traffic conditions, establish the likelihood function Pr(D|H i ), which indicates that in the event of abnormality H i When it occurs, the probability of observing data D; Using Bayes' theorem, calculate the posterior probability Pr(H i |D), that is, after observing data D, the abnormal event H i Probability of occurrence: Among them, Pr(H i |D) means that after observing data D, event H i The posterior probability of occurrence; Pr(D|H i ) represents the hypothetical event H i The probability of the observed data D occurring under the premise of occurrence is called the likelihood function; Pr(H i ) indicates event H i The prior probability of occurrence represents the subjective belief in the occurrence of the event in the absence of any new data D; j Pr(D|H j )·Pr(H j ) represents the combined impact of all possible events, ensuring the normalization of all posterior probabilities so that the sum of the probabilities of all possible events is 1; An event with a higher posterior probability is considered to be more likely to occur; By calculating the posterior probability of each abnormal event in real time, the most likely abnormal event is identified.

10. The heavy-load traffic abnormal event detection and state identification system according to claim 1, characterized in that: The transfer learning unit applies the abnormal event detection model trained under normal weather conditions to adverse weather conditions through the transfer learning method, and fine-tunes and optimizes the model under the new weather conditions to ensure that the model can efficiently and accurately detect abnormal events under various weather conditions; through transfer learning, the deployment process of the abnormal event detection model is accelerated, the time and resource consumption of retraining the model are reduced, and the flexibility and adaptability of the system are improved.

Images