Macroscopic traffic safety analysis method for improving scale effect

By using a Bayesian spatial multi-scale model, the scaling effect is improved by utilizing the correlation between structured and unstructured spaces. This solves the problem of inconsistent statistical results caused by data aggregation in macroscopic traffic safety analysis, and improves the accuracy and portability of accident prediction models.

CN119004958APending Publication Date: 2024-11-22ZHENGZHOU UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411013179.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In macro-level traffic safety analysis, the acquisition of data at different spatial scales leads to scale effects during data aggregation, resulting in inconsistent statistical results and information loss, which affects the accuracy of accident prediction models.

Method used

A Bayesian spatial multiscale model is adopted, which introduces structured and unstructured spatial correlation as a shared random effect between different levels to transmit spatial correlations from higher levels to lower levels. Structured and unstructured spatial correlation terms are used to simulate changes in spatial risk of accidents.

Benefits of technology

It improves the problem of plastic area units caused by scale effects, enhances the accuracy and portability of accident prediction models, and reduces model complexity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004958A_ABST
    Figure CN119004958A_ABST
Patent Text Reader

Abstract

The invention discloses a macroscopic traffic safety analysis method for improving a scale effect, and provides a Bayesian multi-scale joint analysis model for scale effect improvement and a region boundary accident aggregation method based on a data amplification iterative algorithm. The influence of a plastic area unit problem on macroscopic traffic safety research is tried to be solved; in a proposed model framework, scale effects associated with the aggregation are described by sharing spatial correlation items (different combinations of structured and unstructured spatial correlation items between different scale division schemes), and correlation errors between different scales are structured by using a modeling method. Therefore, the scale effect problem in the plastic area unit problem is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of road traffic safety, and in particular to a macroscopic traffic safety analysis method for improving scale effect. Background Art

[0002] In macro traffic safety analysis, the collection and processing of traffic data leads to a multi-level structure of data. Since the data is obtained at different spatial scales, it is necessary to process the data to a unified scale before analyzing it when establishing an accident prediction model. When the data is aggregated from a smaller scale (such as block groups) to a larger geographic scale (such as census tracts), scale effects will occur;

[0003] The scale effect is one aspect of the problem of plastic area units, which refers to the problem of inconsistent statistical results caused by the use of zoning schemes of different scales. On the one hand, this data aggregation may lead to the loss of information at the large scale level, resulting in the influence of some important factors offsetting each other at the large scale level. On the other hand, the aggregation of data reduces the heterogeneity of the spatial distribution of data to a certain extent, resulting in different spatial distributions of traffic accidents at different scales;

[0004] In view of this, the present application provides a macroscopic traffic safety analysis method that improves the scale effect. Summary of the invention

[0005] The present application provides a macroscopic traffic safety analysis method for improving scale effect. The present scheme establishes a Bayesian spatial multi-scale model to improve the scale effect problem. In the proposed model framework, the scale effect associated with such aggregation is described by sharing spatial correlation terms (different combinations of structured and unstructured spatial correlation terms between different scale zoning schemes).

[0006] A macroscopic traffic safety analysis method for improving scale effect is characterized by collecting model hypothesis data based on two zoning schemes of different scales in the same area;

[0007] By introducing structural and non-structural spatial correlations as shared random effects between different levels, the spatial correlation relationships of high levels are transferred to lower levels;

[0008] Structured spatial correlation items and unstructured spatial correlation items are used to simulate the spatial risk variation of accidents;

[0009] The structured spatial correlation term assumes a spatial correlation matrix between all regions to account for the correlation between neighboring regions;

[0010] The unstructured spatial correlation term assumes that the spatial trend of a random normal distribution varies across all regions to account for correlation across spatial units.

[0011] In summary, this application has the following beneficial effects:

[0012] This proposal proposes a Bayesian multi-scale joint analysis model for improving scale effects, attempting to weaken the impact of the plastic area unit problem on macro traffic safety research, and uses modeling methods to structure the errors in spatial correlation between different scales, thereby improving the scale effect problem in the plastic area unit problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a schematic diagram for comparing the structures of various models in this application;

[0014] Figure 2 This is a schematic diagram comparing the results of each model in this application;

[0015] Figure 3 This is a schematic diagram comparing the goodness of fit of each model in this application;

[0016] Figure 4 This is a schematic diagram of the estimation results of the parameters of each model in this application;

[0017] Figure 5 Schematic diagram of the estimation results of the parameters of each model in this application (continued). DETAILED DESCRIPTION

[0018] The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0019] In macro traffic safety analysis, the collection and processing of traffic data leads to a multi-level structure of data. Since the data is obtained at different spatial scales, it is necessary to process the data to a unified scale before analyzing it when establishing an accident prediction model. When the data is aggregated from a smaller scale (such as block groups) to a larger geographic scale (such as census tracts), scale effects will occur;

[0020] The scale effect is one aspect of the problem of plastic area units, which refers to the problem of inconsistent statistical results caused by the use of zoning schemes of different scales. On the one hand, this data aggregation may lead to the loss of information at the large scale level, resulting in the influence of some important factors offsetting each other at the large scale level. On the other hand, the aggregation of data reduces the heterogeneity of the spatial distribution of data to a certain extent, resulting in different spatial distributions of traffic accidents at different scales;

[0021] Accident prediction model is the main tool for macro traffic safety analysis. The selection of different zoning schemes for domain traffic safety analysis will affect the results of the accident prediction model. Traditional accident prediction models, including the commonly used negative binomial and Poisson lognormal models, are based on the same basic assumption that observations should be independent of each other. Because of the hierarchical data structure generated during data aggregation and clustering, this "independence" assumption is difficult to hold in reality. To this end, this application proposes a multi-scale joint analysis model to explain the interconnections between different levels of regional scales.

[0022] 1.1 Model Structure

[0023] This application proposes six different multiscale models that consider spatial correlation to explain scale effects. Spatial correlation is an important consideration in macro safety analysis. Considering spatial correlation between accidents has two main benefits. First, it helps to consider the impact of surrounding geographic units and thus improve model estimation. Second, spatial correlation can be used as a substitute for unobserved variables, assuming that these covariates vary smoothly within the study area. The model assumes that the data are collected based on two different scale zoning schemes in the same area, and each cell in the large-scale zoning scheme is aggregated from several sub-cells in the small-scale zoning scheme. By introducing structural and unstructured spatial correlation as shared random effects between different scales, high-level spatial correlation relationships are transferred to low levels. The structured spatial correlation term considers the correlation between adjacent regions by assuming a spatial correlation matrix between all regions. The unstructured spatial correlation term assumes that the spatial trend of the random normal distribution varies in all regions. Therefore, it is used to represent the correlation within the entire spatial unit rather than the correlation between adjacent regions, and is used to replace part of the information that the results vary with geographical location.

[0024] The following describes how to use the full Bayesian estimation method to build a multi-scale joint analysis model for two-level frequency data.

[0025] 1.1.1 Model Structure

[0026] (1) Model 1

[0027] Model 1 establishes two independent accident prediction models at the cell and sub-cell levels. This model does not introduce correlation terms between scales. The model expression can be written as:

[0028] Yi~Poisson(λi)

[0029] ln(λ i )=a1+β1 ln(Evi)+βx1+v i +u i (1-1)

[0030] Yj~Poisson(λj)

[0031] ln(λ j )=a2+β2ln(Evj)+βx2+v j +u j (1-2)

[0032] Among them, Y i , Y j is the number of accidents at the cell and sub-cell level, λi, λj are the expected values ​​of accidents at the cell and sub-cell level, assuming that the number of accidents follows a Poisson distribution. i ,λ j Logarithmic processing is performed to establish a linear relationship with the explanatory variables, where a1, a2, E vi , E vj , x1, x2, β1, β2 represent the intercept parameter, accident chance variable, other explanatory variables and variable parameters at the cell and sub-cell level respectively; assuming that the unstructured spatial correlation term v i and v j obeys the normal distribution, while the structured spatial correlation term u i and u j is subject to conditional autoregressive distribution;

[0033] Among them, other explanatory variables may vary according to different data sources. Commonly used ones include median household income, intersection density, length of speed limit sections, road network density, etc., and no excessive restrictions are made here; the variable parameter is the coefficient of the variable, which indicates the degree of influence of the variable.

[0034] (2) Model 2

[0035] In model 2, a joint model is established at the cell and sub-cell levels. An unstructured shared spatial association term v is introduced in the sub-cell model. i To establish a connection between the two scales. i Defined in the large-scale model, it is used to characterize the unstructured spatial association between large-scale cells and the unstructured shared spatial association term v of small-scale cells belonging to the same large-scale cell i are the same. The model expression is:

[0036] Yi~Poisson(λi)

[0037] ln(λ i )=a1+β1ln(E vi )+βx1+v i +u i (1-3)

[0038] Y j ~Poisson(λj )

[0039] ln(λ j )=a2+β2ln(E vj )+βx2+v i +u j (1-4)

[0040] (3) Model 3

[0041] In Model 3, a structured shared spatial association term u is introduced in the sub-cell model. i To establish the connection between the two scales. Each sub-cell shares the effect u i , inheriting the structural spatial characteristics of the corresponding large-scale zoning scheme. For cell i and sub-cell j, the model expression

[0042] The formula is:

[0043] Yi~Poisson(λi)

[0044] ln(λi)=a1+β1 ln(Evi)+βx1+v i +u i (1-5)

[0045] Yj~Poisson(λj)

[0046] ln(λj)=a2+β2ln(Evj)+βx2+v j +u i (1-6)

[0047] (4) Model 4

[0048] Model 4 is based on Model 2 and adds the unstructured spatial effect of sub-districts to the model of small-scale zoning schemes. j , v j The unstructured spatial effects at the sub-district level can be captured as a supplement to the random effects in the large-scale zoning scheme model. The model expression is:

[0049] Yi~Poisson(λi)

[0050] ln(λi)=a1+β1 ln(Evi)+βx1+v i +u i (1-7)

[0051] Yj~Poisson(λj)

[0052] ln(λj)=a2+β2ln(Evj)+βx2+v i +v j +uj (1-8)

[0053] (5) Model 5

[0054] Model 5 is based on Model 3 and adds the structured spatial effect u of the sub-district to the model based on the small-scale zoning scheme. j .u j The structured spatial effects at the sub-district level can be captured as a supplement to the random effects in the large-scale zoning scheme model. The model expression is:

[0055] Yi~Poisson(λi)

[0056] ln(λi)=a1+β1 ln(Evi)+βx1+v i +u i (1-9)

[0057] Yj~Poisson(λj)

[0058] ln(λj)=a2+β2ln(Evj)+βx2+v j +u i +u j (1-10)

[0059] (6) Model 6

[0060] In Model 6, the model based on the small-scale zoning scheme includes both scales.

[0061] Structured space and unstructured space effects. The model expression is:

[0062] Yi~Poisson(λi)

[0063] ln(λi)=a1+β1 ln(Evi)+βx1+v i +u i (1-11)

[0064] Yj~Poisson(λj)

[0065] ln(λj)=a2+β2ln(Evj)+βx2+v i +v j +u i +u j (1-12)

[0066] In summary, Model 1 is an independent model that fits accident frequencies at two scales. Models 2 and 4 study the shared unstructured spatial random effects between the two scales, and Models 3 and 5 study the shared structured spatial random effects between the two scales. Model 6 considers both the unstructured random effects and the structured random effects between the two scales. Figure 1 This is a comparison chart of the structures of each model.

[0067] 1.1.2 Model Evaluation Indicators

[0068] The Bayesian Deviation Information Criterion (DIC), Mean Absolute Deviation (MAD) and Mean Squared Prediction Error (MSPE) were used to scientifically evaluate the selected models.

[0069] Similar to Akaike Information Criterion (AIC), DIC takes into account both model fit and complexity. Its calculation method is as follows:

[0070]

[0071] Where: Used to measure the goodness of model fit; p D DIC is the number of effective parameters, which is used to measure the complexity of the model. Generally speaking, the smaller the DIC value, the better the model fit. If the DIC difference is greater than 5, it indicates that there is a significant difference in the goodness of fit between the models.

[0072] MAD and MSPE are calculated as follows:

[0073]

[0074]

[0075] In the formula, The predicted value of the accident in the ith community. The smaller the MAD and MSPE values, the more accurate the model prediction.

[0076] 1.2 Example Analysis

[0077] 1.2.1 Data Collection and Preprocessing

[0078] Data from Hillsborough County, Florida, including crash data, road traffic characteristics, and socio-economic-demographic characteristics, are summarized by census unit group and census tract.

[0079] 1.2.2 Model estimation method

[0080] The statistical software WinBUGS is used to estimate the model. The model solution will adopt the full Bayesian method. Through the Markov chain Monte Carlo simulation sampling process, the distribution and significance level of the model parameters are obtained, and the regional traffic safety evaluation equation is further obtained for the analysis of the model results. In the model estimation, 300,000 cycles are set first. After the convergence is determined, 10,000 cycles are performed to infer the posterior distribution of the parameters. The convergence of the model is monitored by the Brooks-Gelman-Rubin statistic, the visual test of the Markov chain Monte Carlo chain, and the ratio of the montecarlo error to the standard deviation of the estimated value. According to experience, a ratio less than 0.05 indicates that the model results converge.

[0081] 1.2.3 Model results and analysis

[0082] In this section, the model performances of different multiscale joint analysis models are compared, the parameter estimation results are discussed, and finally the results for high accident incidence areas are presented and compared.

[0083] 1.2.3.1 Model Performance Comparison

[0084] Figure 2 and Figure 3 This is a schematic diagram of model evaluation and comparison results, where the calculation time is used to detect the convergence speed of the model. Figure 4 and Figure 5 Schematic diagram of the results of parameter estimation for each model.

[0085] By comparing the results of Model 1, Model 3 and Model 5, the structured shared spatial correlation effect was analyzed. The variance of the structured random effect in Model 5 (sduj: 0.04) was much smaller than that in Model 1 (sduj: 0.53), and the variance of the structured shared random effect in Model 5 (sdui: 0.56) was also higher than that in Model 1 (sdui: 0.47). This was because part of the variability in Model 5 was represented by the shared random effect (sdui: 0.56), indicating that the spatial structure of the census area can reflect the spatial correlation at the level of the census unit group to a certain extent. The results of Model 3 further proved this point. Model 3 is similar to Model 5, except that the model does not include the structured spatial correlation term at the level of the small-scale community. The model performance of Model 3 and Model 5 is comparable, but the convergence speed of Model 3 is much faster than that of other models, which shows that including the structural spatial correlation term at the level of the census area can not only well represent the spatial distribution characteristics of accidents at the level of the census unit group, but also reduce the complexity of the model. Since the model estimation considering spatial correlation is time-consuming, especially for small-scale zoning schemes with a large number of communities and more complex spatial correlations, Model 3 is suitable for situations where more ideal results need to be obtained and model calculation time needs to be saved.

[0086] The unstructured shared spatial correlation effect was tested by comparing the results of Model 1, Model 2, and Model 4. The variability of the unstructured random effects of Model 1 and Model 4 was similar (SDVJ: 0.55, 0.56), but the variability of the structured random effects of Model 4 (SDUJ: 0.19) was much smaller than that of Model 1 (SDUJ: 0.53). This was because part of the variability in Model 4 was represented by the shared random effects (SDVI: 0.48), and it was higher than the variance in Model 1 (SDVI: 0.27). These results indicate that the unstructured spatial correlation at the census tract level cannot well capture the unstructured spatial correlation of census unit groups, but can complement the structured spatial correlation at the census unit group level. Similar to Model 3, in Model 2, replacing the random spatial correlation term of census unit groups with census tracts can reduce model complexity and improve the performance of the model based on census unit groups, based on the shorter convergence time and lower DIC values.

[0087] By comparing the goodness of fit of the models, we can find that: First, at the census tract level, the DIC of the multiscale joint analysis model considering shared random effects (models 2-6) is significantly lower than that of the traditional model (model 1), indicating that adding shared random effects to the model will significantly improve the performance of the census tract-based model, because the spatial correlation terms at the census tract level are also corrected at the census unit group level in the form of shared random effect parameters, effectively compensating for the information loss caused by data aggregation. Second, based on the overall DIC, the models with multiscale shared unstructured correlation terms (models 2 and 4) perform better than the models containing shared structured spatial correlation terms (models 3 and 5), which is consistent with the previous results of modeling at the single spatial unit level.

[0088] 1.2.3.2 Portability Analysis

[0089] On this basis, the model prediction results of census unit groups are aggregated to census districts to infer the safety status at the census district level, which is used to study the portability of the model results. Figure 3 It can be seen that the MAD and MSPE values ​​of the multi-scale model are smaller, indicating that the portability of the multi-scale model results is better than that of the independent model (Model 1). Among them, the portability results of the model with structural shared random effects are better than those of the model with non-structural shared random effects. This is because non-structural random effects can provide more general spatial correlation relationships and are not limited to spatial adjacent structures at different scales. The proposed method can be used to predict safety levels when data at certain scales are missing.

[0090] 1.3 Summary

[0091] Due to the existence of scale effect, the aggregation of data from smaller spatial units to larger spatial units will lead to information loss, resulting in biased parameter estimation and erroneous inference. The main purpose of this chapter is to propose a Bayesian multiscale model to weaken the scale effect problem. In the proposed model framework, the association between different scales is represented by sharing spatial correlation terms, that is, different combinations of structured and unstructured spatial correlation terms between different scales. The performance of the proposed model is evaluated using accident data from two geographic units, namely, census unit groups and census tracts in Hillsborough County, Florida. The results show that the multiscale joint analysis model is an effective accident prediction modeling method, which can obtain more accurate risk estimation results and help study the portability of results at different scales.

[0092] Although the present invention has been illustrated and described with respect to the preferred embodiments, it will be appreciated by those skilled in the art that various changes and modifications may be made to the present invention without departing from the scope of the present invention as defined by the claims.

Claims

1. A macroscopic traffic safety analysis method for improving scale effect, characterized in that: Collecting model assumptions based on two different scale zoning schemes for the same area; By introducing structural and non-structural spatial correlations as shared random effects between different levels, the spatial correlation relationships of high levels are transferred to lower levels; Structured spatial correlation items and unstructured spatial correlation items are used to simulate the spatial risk variation of accidents; The structured spatial correlation term assumes a spatial correlation matrix between all regions to account for the correlation between neighboring regions; The unstructured spatial correlation term assumes that the spatial trend of a random normal distribution varies across all regions to account for correlation across spatial units.

2. A macroscopic traffic safety analysis method for improving scale effect according to claim 1, characterized in that: The two zoning schemes of different scales include a large-scale zoning scheme and a small-scale zoning scheme. Each cell in the large-scale zoning scheme is aggregated from a number of sub-cells in the small-scale zoning scheme.

3. A macroscopic traffic safety analysis method for improving scale effect according to claim 2, characterized in that: Constructing Model 1: Two independent accident prediction models are established at the cell and sub-cell levels. The expression of Model 1 is: ln(λj)=a2+β2ln(Evj)+βx2+v j +u j (1-2) Among them, Y i , Y j is the number of accidents at the cell and sub-cell level, λi, λj are the expected values ​​of accidents at the cell and sub-cell level; a1, a2, E vi , E vj , x1, x2, β1, β2 represent the intercept parameter, accident chance variable, other explanatory variables and variable parameters at the community and sub-community level respectively; v i and v j is an unstructured spatial correlation term and obeys normal distribution, u i and u j It is a structured spatial correlation term and obeys the conditional autoregressive distribution.

4. The macroscopic traffic safety analysis method for improving scale effect according to claim 2 is characterized in that: Model 2: A joint model is built at the cell and sub-cell levels. An unstructured shared spatial association term v is introduced in the sub-cell model. i To establish the connection between the two scales, the expression of Model 2 is: ln(λj)=a2+β2ln(Evj)+βx2+v i +u j 。 (1-4) 5. The macroscopic traffic safety analysis method for improving scale effect according to claim 2 is characterized in that: Model 3: By introducing a structured shared spatial association term u in the sub-cell model i To establish the connection between the two scales, each sub-cell shares the effect u i , inheriting the structural spatial characteristics of the corresponding large-scale zoning scheme, for cell i and sub-cell j, the model expression is: ln(λj)=a2+β2ln(Evj)+βx2+v j +u i 。 (1-6) 6. A macroscopic traffic safety analysis method for improving scale effect according to claim 4, characterized in that: Constructing Model 4: Based on Model 2, Model 4 adds the unstructured spatial effect of sub-districts to the model of small-scale zoning schemes. j , v j The unstructured spatial effect at the sub-district level can be captured as a supplement to the random effect in the large-scale zoning scheme model. The model expression is: ln(λj)=a2+β2 ln(Evj)+βx2+v i +v j +u j 。 (1-8) 7. The macroscopic traffic safety analysis method for improving scale effect according to claim 5 is characterized in that: Constructing Model 5: Based on Model 3, Model 5 adds the structured spatial effect u of sub-districts to the model based on the small-scale zoning scheme. j ,u j The structured spatial effect at the sub-district level can be captured as a supplement to the random effect in the large-scale zoning scheme model. The model expression is: ln(λj)=a2+β2 ln(Evj)+βx2+v j +u i +u j 。 (1-10) 8. The macroscopic traffic safety analysis method for improving scale effect according to claim 2 is characterized in that: Model 6: The model based on the small-scale zoning scheme includes both structured and unstructured spatial effects at two scales. The model expression is: ln(λj)=a2+β2ln(Evj)+βx2+v i +v j +you i +you j (1-12)