Method and system for detecting vehicle collision risk based on hox process

Abstract

The invention discloses a method and system for detecting a vehicle collision risk based on a hox process, and relates to the technical field of vehicle collision risk detection, and the method comprises the steps: obtaining vehicle collision data; modeling the vehicle collision data into a hokes process model combining excitation and inhibition effects; constructing a Gibbs sampler, sampling model parameters of the Horkes process model through the Gibbs sampler, and updating the Horkes process model parameters; and predicting the time interval of the collision of the next vehicle based on the updated hokes process model, if the collision point of the next real vehicle is in the time interval, indicating that the collision risk of the vehicle is not changed, otherwise, indicating that the collision risk of the vehicle is changed. According to the invention, the change position of the vehicle collision risk in a specific area and a specific time period can be simply and effectively detected, and an accurate risk assessment tool is provided for a traffic management department.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle collision risk detection, and more particularly to a method and system for detecting vehicle collision risk based on Hawkes process. Background Art

[0002] Vehicle collisions are a major hidden danger to public safety. Accurately predicting future vehicle collision trends can effectively optimize traffic management and resource allocation and improve road safety. How to effectively judge the changes in collision risks in a specific area at different times based on past collision times is a highly challenging problem. Historical vehicle collision data are usually presented in the form of spatiotemporal event points. For example, a collision event can be described by the specific time and place where it occurred. Point process data are characterized by a series of discrete events that occur over time and have a wide range of applications in many fields such as finance, neuroscience, and social networks. As a subclass of point processes, Hawkes processes have attracted attention for their ability to model self-excitation and clustering behaviors.

[0003] However, traditional Hawkes process modeling has limitations in detecting changes in vehicle collision risk because it relies on the assumption that the distribution of collision events remains constant over time. In reality, the collision risk of a specific area can change significantly over time due to a variety of factors. For example, factors such as seasonal changes, road construction, or major events can cause a significant increase or decrease in the collision risk of a certain area.

[0004] Therefore, how to accurately capture the changes in collision risks in different time periods based on historical vehicle collision data in a dynamic environment is a technical problem that needs to be urgently solved in current traffic safety management. Summary of the invention

[0005] In view of this, the present invention provides a method and system for detecting vehicle collision risk based on Hawkes process, which can accurately and timely detect the change points in Hawkes process, simply and effectively detect the changing position of vehicle collision risk in specific areas and time periods, and provide accurate risk assessment tools for traffic management departments. This not only helps to identify potential high-risk time areas in advance, but also provides strong decision-making support for vehicle collision risk management, thereby optimizing the allocation of traffic resources and the implementation of intervention measures, and greatly improving the level of road safety.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A method for detecting vehicle collision risk based on Hawkes process, comprising:

[0008] Obtain vehicle collision data;

[0009] The vehicle crash data is modeled as a Hawkes process model that combines excitation and inhibition effects;

[0010] Construct a Gibbs sampler, sample the model parameters of the Hawkes process model through the Gibbs sampler, and update the Hawkes process model parameters;

[0011] The time interval of the next vehicle collision is predicted based on the updated Hawkes process model. If the next real vehicle collision point is within the above time interval, it means that the vehicle collision risk has not changed, otherwise it has changed.

[0012] Preferably, the vehicle collision data is modeled as a Hawkes process model combining excitation and inhibition, specifically including:

[0013] 1) Model the vehicle collision data as a nonlinear Hawkes process, defined as:

[0014]

[0015] in, is the conditional intensity function, is the upper bound of strength, is the sigmoid function, is the baseline activation, where is the field of real numbers, is the influence function, which means arrive The self-excitation effect of the vehicle crash event that occurred, whose sum captures the impact of all past vehicle crash events, will be Modeled as a linear combination of multiple basis functions:

[0016]

[0017] in, is the bth basis function, is the mixing weight, B is the total number of basis functions;

[0018]

[0019] in, and The bth basis function represents the cumulative impact of past events on time t. is determined by the corresponding The vector composed of

[0020] 2) Based on the nonlinear Hawkes process, the probability density function of the Hawkes process model before the change point is obtained:

[0021]

[0022] in, Indicated in and Under the conditions The joint probability density function of the vehicle collision event occurs, assuming exist It is observed in the time interval, M is the total number of vehicle collision data of the Hawkes process without change points, and are Hawkes process model parameters, express time The value of the function, express time The value of the function;

[0023] 3) Set represents the sequence of timestamps generated by the Hawkes process, is the length of vehicle collision data of Hawkes process with change point, for timestamp , defining the most recent change point index as , and assuming that the timestamps before and after the change point are independent of each other, the timestamp after the change point The probability density function of is expressed as:

[0024]

[0025] in, Indicated in and Under the conditions The probability of occurrence within a time interval, express time The value of the function;

[0026] 4) Based on the timestamp after the change point The probability density function of the model is used to calculate the posterior value of the model parameters. :

[0027]

[0028] in, The prior is Gaussian , is a Gaussian distribution, is the covariance matrix, The prior is an uninformative prior ;

[0029] 5) Calculate the predicted distribution of the next vehicle collision time using the posterior values ​​of the model parameters:

[0030]

[0031] in, Observed data points for vehicle collision The timestamp of the next vehicle collision The probability distribution of For Condition, the next vehicle collision timestamp The conditional probability distribution of .

[0032] Preferably, constructing a Gibbs sampler specifically includes:

[0033] 1) Timestamp after the change point Introducing two sets of latent variables, Pólya-Gamma variables and labeled Poisson process, into the probability density function of , we get the augmented likelihood :

[0034]

[0035] in, yes Each on The vector formed, It is an observation point Pólya-Gamma variables on , , is the Pólya-Gamma distribution, where , , express Where Function value, express Where Function value, where is the Pólya-Gamma variable marking the points of the Poisson process; Representation interval There is a realization of the Poisson process with a probability measure of ;

[0036] 2) Combining the augmented likelihood and prior, we get the augmented joint distribution:

[0037]

[0038] 3) Get the Gibbs sampler based on the joint distribution:

[0039]

[0040]

[0041]

[0042]

[0043] in, is given , hour The conditional probability distribution of Indicates a given hour Pólya-Gamma conditional probability distribution of; is given The intensity function of the underlying Poisson process is Indicates a given hour Pólya-Gamma conditional probability distribution of; is given hour The conditional probability distribution of Indicates a given hour Gamma conditional probability distribution, , represents the number of points in the Poisson process. ; is given of Conditional probability distribution, ,in is a diagonal matrix, The elements are , the next The elements are , ; ,in is a vector, forward The elements are , the next The elements are .

[0044] Preferably, the parameters of the Hawkes process model are obtained by iterative sampling using a Gibbs sampler to obtain a series of samples from the posterior distribution of the model parameters. ;

[0045] The model parameters of the inhibiting Hawkes process model are sampled by a Gibbs sampler, and the Hawkes model parameters are updated according to the obtained model parameters.

[0046] A system for detecting vehicle collision risk based on Hawkes process, comprising:

[0047] Data acquisition module: obtain vehicle collision data;

[0048] Hawkes process model building module: Model vehicle collision data as a Hawkes process model that combines excitation and inhibition;

[0049] Gibbs sampler construction and parameter update module: construct a Gibbs sampler, sample the model parameters of the Hawkes process model through the Gibbs sampler, and update the Hawkes process model parameters;

[0050] Vehicle collision risk prediction module: Based on the updated Hawkes process model, the time interval of the next vehicle collision is predicted. If the next real vehicle collision point is in the above time interval, it means that the vehicle collision risk has not changed, otherwise it has changed.

[0051] Preferably, the vehicle collision data is modeled as a Hawkes process model combining excitation and inhibition, specifically including:

[0052] 1) Model the vehicle collision data as a nonlinear Hawkes process, defined as:

[0053]

[0054] in, is the conditional intensity function, is the upper bound of strength, is the sigmoid function, is the baseline activation, where is the field of real numbers, is the influence function, which means arrive The self-excitation effect of the vehicle crash event that occurred, whose sum captures the impact of all past vehicle crash events, will be Modeled as a linear combination of multiple basis functions:

[0055]

[0056] in, is the bth basis function, is the mixing weight, B is the total number of basis functions;

[0057]

[0058] in, and The bth basis function represents the cumulative impact of past events on time t. is determined by the corresponding The vector composed of

[0059] 2) Based on the nonlinear Hawkes process, the probability density function of the Hawkes process model before the change point is obtained:

[0060]

[0061] in, Indicated in and Under the conditions The joint probability density function of the vehicle collision event occurs, assuming exist It is observed in the time interval, M is the total number of vehicle collision data of the Hawkes process without change points, and are Hawkes process model parameters, express time The value of the function, express time The value of the function;

[0062] 3) Set represents the sequence of timestamps generated by the Hawkes process, is the length of vehicle collision data of Hawkes process with change point, for timestamp , defining the most recent change point index as , and assuming that the timestamps before and after the change point are independent of each other, the timestamp after the change point The probability density function of is expressed as:

[0063]

[0064] in, Indicated in and Under the conditions The probability of occurrence within a time interval, express time The value of the function;

[0065] 4) Based on the timestamp after the change point The probability density function of the model is used to calculate the posterior value of the model parameters. :

[0066]

[0067] in, The prior is Gaussian , is a Gaussian distribution, is the covariance matrix, The prior is an uninformative prior ;

[0068] 5) Calculate the predicted distribution of the next vehicle collision time using the posterior values ​​of the model parameters:

[0069]

[0070] in, Observed data points for vehicle collision The timestamp of the next vehicle collision The probability distribution of For Condition, the next vehicle collision timestamp The conditional probability distribution of .

[0071] Preferably, constructing a Gibbs sampler specifically includes:

[0072] 1) Timestamp after the change point Introducing two sets of latent variables, Pólya-Gamma variables and labeled Poisson process, into the probability density function of , we get the augmented likelihood :

[0073]

[0074] in, yes Each on The vector formed, It is an observation point Pólya-Gamma variables on , , is the Pólya-Gamma distribution, where , , express Where Function value, express Where Function value, where is the Pólya-Gamma variable marking the points of the Poisson process; Representation interval There is a realization of the Poisson process with a probability measure of ;

[0075] 2) Combining the augmented likelihood and prior, we get the augmented joint distribution:

[0076]

[0077] 3) Get the Gibbs sampler based on the joint distribution:

[0078]

[0079]

[0080]

[0081]

[0082] in, is given , hour The conditional probability distribution of Indicates a given hour Pólya-Gamma conditional probability distribution of; is given The intensity function of the underlying Poisson process is Indicates a given hour Pólya-Gamma conditional probability distribution of; is given hour The conditional probability distribution of Indicates a given hour Gamma conditional probability distribution, , represents the number of points in the Poisson process. ; is given of Conditional probability distribution, ,in is a diagonal matrix, The elements are , the next The elements are , ; ,in is a vector, forward The elements are , the next The elements are .

[0083] Preferably, the parameters of the Hawkes process model are obtained by iterative sampling using a Gibbs sampler to obtain a series of samples from the posterior distribution of the model parameters. ;

[0084] The model parameters of the inhibiting Hawkes process model are sampled by a Gibbs sampler, and the Hawkes model parameters are updated according to the obtained model parameters.

[0085] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for detecting vehicle collision risk based on Hawkes process, which uses data augmentation to solve the non-conjugate reasoning problem of Hawkes process and converts the non-conjugate reasoning problem into a conditional conjugate reasoning problem. This novel method makes the change point detection of Hawkes process more accurate and efficient, and can effectively judge the changes in collision risk at different times in a specific area, which is beneficial for management departments to formulate policies to reduce vehicle collisions. An analytical Gibbs sampler is developed for the proposed model, which realizes closed iterative sampling of model parameters. This simplifies the reasoning process and reduces the computational burden associated with non-conjugate scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0087] Figure 1 A flow chart of a method for detecting vehicle collision risk based on Hawkes process provided by the present invention.

[0088] Figure 2 A block diagram of a system for detecting vehicle collision risk based on Hawkes process provided by the present invention.

[0089] Figure 3 This is an analysis diagram of vehicle collision data results using CoBay-CPD.

[0090] Figure 4 This is an analysis diagram of the vehicle collision data results using SMCPD.

[0091] Figure 5 This is an analysis diagram of the results of vehicle collision data using SVCPD.

[0092] Figure 6 This is an analysis diagram of vehicle collision data results using SVCPD+Inhibiton. DETAILED DESCRIPTION

[0093] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0094] The embodiment of the present invention discloses a method for detecting vehicle collision risk based on Hawkes process, such as Figure 1 As shown, including:

[0095] Obtain vehicle collision data;

[0096] The vehicle crash data is modeled as a Hawkes process model that combines excitation and inhibition effects;

[0097] Construct a Gibbs sampler, sample the model parameters of the Hawkes process model through the Gibbs sampler, and update the Hawkes process model parameters;

[0098] The time interval of the next vehicle collision is predicted based on the updated Hawkes process model. If the next real vehicle collision point is within the above time interval, it means that the vehicle collision risk has not changed, otherwise it has changed.

[0099] Furthermore, the specific implementation process of each step of the present invention is discussed.

[0100] Hawkes process is a method used to model vehicle collision data. Traditional Hawkes process only uses positive influence functions to avoid negative intensities. However, it is more appropriate to model vehicle collision data as a Hawkes process that combines excitation and inhibition. For example, a vehicle collision event may have occurred and been reported ahead of the road. In this case, drivers traveling on the road will be more vigilant, so the occurrence of vehicle collision events may be reduced. This is a manifestation of the inhibitory effect between events in the vehicle collision process. In order to combine excitation and inhibition, many studies have proposed nonlinear Hawkes processes that allow influence functions to be negative. The present invention models vehicle collision data as a nonlinear Hawkes process, defined as:

[0101] (1)

[0102] in, is the conditional intensity function, is the upper bound of strength, is the sigmoid function, is the baseline activation, where is the field of real numbers, is the influence function, which means arrive The self-excitation effect of the vehicle crash event that occurred, whose sum captures the effect of all past vehicle crash events, The strength of the representation depends on the history. Since the sigmoid function existence, and can all be negative, thus enabling it to capture inhibition.

[0103] In order to obtain a flexible influence function that can better characterize vehicle collision data, Modeled as a linear combination of multiple basis functions:

[0104] (2)

[0105] in, is the bth basis function, is the mixed weight, and B is the total number of basis functions. The present invention selects a support interval of The scaled and translated beta distribution of is used as the basis function. Defined with bounded support Rather than unbounded support The basis function is used to assume that events that occurred too early will not affect the current time. For example, the impact of a vehicle collision ten years ago on a vehicle collision now is negligible and can be ignored. This choice ensures that the subsequent calculations are more efficient. This also makes the present invention more efficient and can more quickly determine the position where the vehicle collision direction changes. The expression can be expressed in vector form:

[0106] (3)

[0107] in, and The bth basis function represents the cumulative impact of past events on time t. is determined by the corresponding Therefore, the probability density function of the model proposed in the present invention is:

[0108] (4)

[0109] in, Indicated in and Under the conditions The joint probability density function of the vehicle collision event occurs, assuming exist It is observed in the time interval, M is the total number of vehicle collision data of the Hawkes process without change points, and are Hawkes process model parameters, express time The value of the function, express time The value of the function.

[0110] The above section outlines the Hawkes process without change points, but due to the influence of many factors, such as weather changes, changes in road conditions, etc., the collision risk of vehicles in different time periods will change significantly. Therefore, the following introduces the Hawkes process with change points, which can better model vehicle collision data.

[0111] The present invention uses a Hawkes process with change points and how to design a Bayesian two-step change point detection method to detect these change points, that is, how to detect the time point when the vehicle collision risk changes. In the Bayesian two-step change point detection method, the goal is to determine the change point of the potential dynamic change of the point process. The method is divided into two steps: estimation and prediction. represents the sequence of timestamps generated by the Hawkes process, is the data length of the Hawkes process with change points, i.e., vehicle collision data, and its parameters are ,think May change at some timestamps. , defining the most recent change point index as , and assumes that the timestamps before and after the change point are independent of each other. Under this assumption, in the estimation step, Bayesian two-step change point detection needs to be based on Estimate the posterior of the model parameters. To infer the posterior, the timestamp after the change point is converted into The likelihood of is expressed as:

[0112] (5)

[0113] According to Bayes' theorem, the posterior value of the model parameter is expressed as:

[0114] (6)

[0115] in, The prior is Gaussian , is a Gaussian distribution, is the covariance matrix, The prior is an uninformative prior .exist A Gaussian prior is used because it is equivalent to Regularizer, which can stabilize parameter estimates when observation data are insufficient.

[0116] Then, in the prediction step, the posterior values ​​of the model parameters are used to calculate the predicted distribution of the next timestamp, that is, the predicted distribution of the time when the next vehicle collision will occur is:

[0117] (7)

[0118] in, Observed data points for vehicle collision The timestamp of the next vehicle collision The probability distribution of For Condition, the next vehicle collision timestamp The conditional probability distribution of .

[0119] This formula calculates the vehicle collision observation data points given the previous The timestamp of the next vehicle collision distribution.

[0120] The above is the Hawkes process model designed by the present invention, but in the implementation process, it is challenging to solve formula (6) and formula (7). For formula (6), the non-conjugate nature between the point process likelihood and the prior makes it impossible to obtain an analytical posterior. For formula (7), it is also difficult to find the integral. Therefore, the present invention uses a sampling method for approximation, but traditional sampling algorithms such as MCMC algorithms often lack analytical expressions. Due to their low computational efficiency, they seriously affect the timeliness of change point detection. Therefore, the present invention obtains model parameters through a Gibbs sampler and .

[0121] The present invention adopts a data augmentation strategy, and enhances the likelihood function of the Hawkes process by introducing auxiliary latent variables, so that the enhanced likelihood condition is conjugate to the prior. On the basis of the conditional conjugate model, an analytical Gibbs sampler is derived, which can effectively obtain the posterior sample of the parameter. That is, the present invention can estimate the parameters of the inhibitory Hawkes process just proposed more efficiently and accurately through the data augmentation method, so as to effectively predict the time interval of the next vehicle collision, thereby detecting whether the vehicle collision risk change at the next time point changes. Specifically, the Pólya-Gamma variable and the labeled Poisson process are combined.

[0122] Timestamp after the change point Introducing two sets of latent variables, Pólya-Gamma variables and labeled Poisson process, into the probability density function of , we get the augmented likelihood :

[0123] (8)

[0124] in, yes Each on The vector formed, It is an observation point Pólya-Gamma variables on , , is the Pólya-Gamma distribution, where , , express Where Function value, express Where Function value, where is the Pólya-Gamma variable marking the points of the Poisson process; Representation interval There is a realization of the Poisson process with a probability measure of , the intensity is . Parameters in the enhanced likelihood has a Gaussian form, so it is conjugate to the Gaussian prior.

[0125] Combining the augmented likelihood and prior, we get the augmented joint distribution:

[0126] (9)

[0127] Due to the conditional conjugacy of the augmented joint distribution, the closed-form conditional density of all variables can be derived, which naturally leads to the analytic Gibbs sampler:

[0128] (10)

[0129] (11)

[0130] (12)

[0131] (13)

[0132] in, is given , hour The conditional probability distribution of Indicates a given hour Pólya-Gamma conditional probability distribution of; is given The intensity function of the underlying Poisson process is Indicates a given hour Pólya-Gamma conditional probability distribution of; is given hour The conditional probability distribution of Indicates a given hour Gamma conditional probability distribution, , represents the number of points in the Poisson process. ; is given of Conditional probability distribution, ,in is a diagonal matrix, The elements are , the next The elements are , ; ,in is a vector, forward The elements are , the next The elements are .

[0133] By iteratively sampling using the Gibbs sampler, a series of samples can be obtained from the posterior distribution of the model parameters to obtain the parameters of the Hawkes process model .

[0134] The model parameters of the Hawkes process model are sampled by the Gibbs sampler, and the Hawkes model parameters are updated according to the obtained model parameters, and the interval of the next point is sampled, that is, the time interval of the next vehicle collision. If the point of the next real vehicle collision is within this interval, it means that the vehicle collision risk has not changed, otherwise, it has changed.

[0135] In order to verify the effect of the present invention, the present invention is compared with several baseline methods on two real data sets to determine whether the method of the present invention has better change point detection performance.

[0136] The proposed method (CoBay-CPD) is compared with several Bayesian change point detection (BCPD) methods that aim to address the non-conjugation challenge of Hawkes processes:

[0137] (1) SMCPD combines BCPD and Sequential Monte Carlo (SMC). Similar to the method of the present invention, it is a sampling-based method to solve non-conjugate inference in the BCPD framework.

[0138] (2) SVCPD similarly combines BCPD and Stein variational inference to solve the non-conjugate inference problem in the BCPD framework. The difference is that this method uses variational techniques to infer the posterior of model parameters.

[0139] (3) SVCPD+Inhibition is an extension of SVCPD, which adds a nonlinear Hawkes process with inhibition. This baseline is designed because the original SVCPD only considers the linear Hawkes process.

[0140] Four metrics are used to evaluate the performance of all methods. (1) False Negative Rate (FNR) quantifies the probability that a change point is mistakenly identified as a non-change point and is calculated as: False negative cases are crucial in many applications because they indicate that some key changes have not been noticed. (2) The false positive rate (FPR) quantifies the probability that a stable point is mistakenly identified as a change point and is calculated as It is also important to minimize false positives, as frequent false positives waste resources. (3) The mean square error (MSE) measures the predicted next timestamp ( The distance between the average value and the actual timestamp (calculated as ). This metric evaluates the accuracy of the model in predicting the next data point. (4) The efficiency of the method is measured by the running time (RT).

[0141] A dataset from vehicle collisions in New York City was analyzed, with the following details:

[0142] New York City Vehicle Collision Data: The New York City Vehicle Collision Dataset contains approximately 1.05 million vehicle collision records, each of which contains information about the time and location of the collision. In our experiments, the records for October 14, 2017 were selected.

[0143] Since there are no true change points in real-world data, the change points reported by Zhou et al. in 2020 are used as the true change points in the New York City dataset.

[0144] The experimental results are as follows:

[0145] Table 1: Comparison of FNR, FPR, MSE, and RT of CoBay-CPD with other baselines on vehicle collisions.

[0146]

[0147] Table 1 presents various metric results of the four change-point detection methods on the New York City data. SMCPD has higher FNR and FPR. The high FPR is due to the excessive number of change points detected by SMCPD, while the high FNR is due to the excessive number of change points detected by SMCPD and inaccurate detection. On the other hand, SVCPD exhibits high FNR and low FPR due to detecting too few change points. SVCPD+Suppression achieves relatively balanced FNR and FPR, indicating the beneficial impact of adopting the nonlinear Hawkes process. Compared with all baseline models, CoBay-CPD has the lowest FNR, FPR, MSE, and RT in change-point detection, showing excellent accuracy and efficiency.

[0148] Figure 3-Figure 6 The change point detection results of the four methods on NYC data are shown. It is worth noting that SVCPD detects fewer change points, while SMCPD identifies too many change points. The change points detected by CoBay-CPD are 43, 110, 160, 194, 284, 338, and 398, corresponding to the times of 2:30, 9:00, 12:00, 13:10, 16:00, 17:55, and 20:00, respectively. These timestamps coincide with the rush hour on weekdays.

[0149] In this embodiment, a system for detecting vehicle collision risk based on Hawkes process is disclosed. Figure 2 As shown, including:

[0150] Data acquisition module: obtain vehicle collision data;

[0151] Hawkes process model building module: Model vehicle collision data as a Hawkes process model that combines excitation and inhibition;

[0152] Gibbs sampler construction and parameter update module: construct a Gibbs sampler, sample the model parameters of the Hawkes process model through the Gibbs sampler, and update the Hawkes process model parameters;

[0153] Vehicle collision risk prediction module: Based on the updated Hawkes process model, the time interval of the next vehicle collision is predicted. If the next real vehicle collision point is in the above time interval, it means that the vehicle collision risk has not changed, otherwise it has changed.

[0154] The specific implementation process of each module is the same as the method implementation method, which will not be repeated here.

[0155] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0156] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting vehicle collision risk based on Hawkes process, characterized in that: include: Obtain vehicle collision data; The vehicle crash data is modeled as a Hawkes process model that combines excitation and inhibition effects; Construct a Gibbs sampler, sample the model parameters of the Hawkes process model through the Gibbs sampler, and update the Hawkes process model parameters; The time interval of the next vehicle collision is predicted based on the updated Hawkes process model. If the next real vehicle collision point is within the above time interval, it means that the vehicle collision risk has not changed, otherwise it has changed.

2. The method for detecting vehicle collision risk based on Hawkes process according to claim 1, characterized in that: The vehicle collision data is modeled as a Hawkes process model that combines excitation and inhibition, including: 1) Model the vehicle collision data as a nonlinear Hawkes process, defined as: ； in, is the conditional intensity function, is the upper bound of strength, is the sigmoid function, is the baseline activation, where is the field of real numbers, is the influence function, which means arrive The self-excitation effect of the vehicle crash event that occurred, whose sum captures the impact of all past vehicle crash events, will be Modeled as a linear combination of multiple basis functions: ； in, is the bth basis function, is the mixing weight, B is the total number of basis functions; ； in, and The bth basis function represents the cumulative impact of past events on time t. is determined by the corresponding The vector composed of 2) Based on the nonlinear Hawkes process, the probability density function of the Hawkes process model before the change point is obtained: ； in, Indicated in and Under the conditions The joint probability density function of the vehicle collision event occurs, assuming exist It is observed in the time interval, M is the total number of vehicle collision data of the Hawkes process without change points, and are Hawkes process model parameters, express time The value of the function, express time The value of the function; 3) Set represents the sequence of timestamps generated by the Hawkes process, is the length of vehicle collision data of Hawkes process with change point, for timestamp , defining the most recent change point index as , and assuming that the timestamps before and after the change point are independent of each other, the timestamp after the change point The probability density function of is expressed as: ； in, Indicated in and Under the conditions The probability of occurrence within a time interval, express time The value of the function; 4) Based on the timestamp after the change point The probability density function of the model is used to calculate the posterior value of the model parameters. : ； in, The prior is Gaussian , is a Gaussian distribution, is the covariance matrix, The prior is an uninformative prior ; 5) Calculate the predicted distribution of the next vehicle collision time using the posterior values ​​of the model parameters: ； in, Observed data points for vehicle collision The timestamp of the next vehicle collision The probability distribution of For Condition, the next vehicle collision timestamp The conditional probability distribution of .

3. The method for detecting vehicle collision risk based on Hawkes process according to claim 2, characterized in that: Construct a Gibbs sampler, including: 1) Timestamp after the change point Introducing two sets of latent variables, Pólya-Gamma variables and labeled Poisson process, into the probability density function of , we get the augmented likelihood : ； in, yes Each on The vector formed, It is an observation point Pólya-Gamma variables on , , is the Pólya-Gamma distribution, where , , express Where Function value, express Where Function value, where is the Pólya-Gamma variable marking the points of the Poisson process; Representation interval There is a realization of the Poisson process with a probability measure of ; 2) Combining the augmented likelihood and prior, we get the augmented joint distribution: ； 3) Get the Gibbs sampler based on the joint distribution: ； ； ； ； in, is given , hour The conditional probability distribution of Indicates a given hour Pólya-Gamma conditional probability distribution of; Indicates a given The intensity function of the underlying Poisson process is Indicates a given hour Pólya-Gamma conditional probability distribution of; Indicates a given hour The conditional probability distribution of Indicates a given hour Gamma conditional probability distribution, , represents the number of points in the Poisson process. ; is given of Conditional probability distribution, ,in is a diagonal matrix, The elements are , the next The elements are , ; ,in is a vector, forward The elements are , the next The elements are .

4. The method for detecting vehicle collision risk based on Hawkes process according to claim 3, characterized in that: Iterative sampling is performed through the Gibbs sampler to obtain a series of samples from the posterior distribution of the model parameters, thereby obtaining the parameters of the Hawkes process model ; The model parameters of the inhibiting Hawkes process model are sampled by a Gibbs sampler, and the Hawkes model parameters are updated according to the obtained model parameters.

5. A system for detecting vehicle collision risk based on Hawkes process, characterized in that: include: Data acquisition module: obtain vehicle collision data; Hawkes process model building module: Model vehicle collision data as a Hawkes process model that combines excitation and inhibition; Gibbs sampler construction and parameter update module: construct a Gibbs sampler, sample the model parameters of the Hawkes process model through the Gibbs sampler, and update the Hawkes process model parameters; Vehicle collision risk prediction module: Based on the updated Hawkes process model, the time interval of the next vehicle collision is predicted. If the next real vehicle collision point is in the above time interval, it means that the vehicle collision risk has not changed, otherwise it has changed.

6. The system for detecting vehicle collision risk based on Hawkes process according to claim 5, characterized in that: The vehicle collision data is modeled as a Hawkes process model that combines excitation and inhibition, including: 1) Model the vehicle collision data as a nonlinear Hawkes process, defined as: ； in, is the conditional intensity function, is the upper bound of strength, is the sigmoid function, is the baseline activation, where is the field of real numbers, is the influence function, which means arrive The self-excitation effect of the vehicle crash event that occurred, whose sum captures the impact of all past vehicle crash events, will be Modeled as a linear combination of multiple basis functions: ； in, is the bth basis function, is the mixing weight, B is the total number of basis functions; ； in, and The bth basis function represents the cumulative impact of past events on time t. is determined by the corresponding The vector composed of 2) Based on the nonlinear Hawkes process, the probability density function of the Hawkes process model before the change point is obtained: ； in, Indicated in and Under the conditions The joint probability density function of the vehicle collision event occurs, assuming exist It is observed in the time interval, M is the total number of vehicle collision data of the Hawkes process without change points, and are Hawkes process model parameters, express time The value of the function, express time The value of the function; 3) Set represents the sequence of timestamps generated by the Hawkes process, is the length of vehicle collision data of Hawkes process with change point, for timestamp , defining the most recent change point index as , and assuming that the timestamps before and after the change point are independent of each other, the timestamp after the change point The probability density function of is expressed as: ； in, Indicated in and Under the conditions The probability of occurrence within a time interval, express time The value of the function; 4) Based on the timestamp after the change point The probability density function of the model is used to calculate the posterior value of the model parameters. : ； in, The prior is Gaussian , is a Gaussian distribution, is the covariance matrix, The prior is an uninformative prior ; 5) Calculate the predicted distribution of the next vehicle collision time using the posterior values ​​of the model parameters: ； in, Observed data points for vehicle collision The timestamp of the next vehicle collision The probability distribution of For Condition, the next vehicle collision timestamp The conditional probability distribution of .

7. The system for detecting vehicle collision risk based on Hawkes process according to claim 6, characterized in that: Construct a Gibbs sampler, including: 1) Timestamp after the change point Introducing two sets of latent variables, Pólya-Gamma variables and labeled Poisson process, into the probability density function of , we get the augmented likelihood : ； in, yes Each on The vector formed, It is an observation point Pólya-Gamma variables on , , is the Pólya-Gamma distribution, where , , express Where Function value, express Where Function value, where is the Pólya-Gamma variable marking the points of the Poisson process; Representation interval There is a realization of the Poisson process with a probability measure of ; 2) Combining the augmented likelihood and prior, we get the augmented joint distribution: ； 3) Get the Gibbs sampler based on the joint distribution: ； ； ； ； in, is given , hour The conditional probability distribution of Indicates a given hour Pólya-Gamma conditional probability distribution of; is given The intensity function of the underlying Poisson process is Indicates a given hour Pólya-Gamma conditional probability distribution of; is given hour The conditional probability distribution of Indicates a given hour Gamma conditional probability distribution, , represents the number of points in the Poisson process. ; is given of Conditional probability distribution, ,in is a diagonal matrix, The elements are , the next The elements are , ; ,in is a vector, forward The elements are , the next The elements are .

8. The system for detecting vehicle collision risk based on Hawkes process according to claim 7, characterized in that: Iterative sampling is performed through the Gibbs sampler to obtain a series of samples from the posterior distribution of the model parameters, thereby obtaining the parameters of the Hawkes process model ; The model parameters of the inhibiting Hawkes process model are sampled by a Gibbs sampler, and the Hawkes model parameters are updated according to the obtained model parameters.

Images