Signal intersection vehicle delay distribution estimation method

Abstract

The invention relates to a signal intersection vehicle delay distribution estimation method, which is characterized in that signal intersection vehicle delay distribution estimation is carried out based on network connection vehicle trajectory data by using a vehicle delay distribution model, and the vehicle delay distribution model comprises a convolutional neural network and a mixed density network. The method comprises the following steps: constructing a network-connected vehicle position matrix based on network-connected vehicle trajectory data, and stacking the network-connected vehicle position matrix to obtain input features; inputting the input features into the convolutional neural network to extract feature data; based on the feature data, the distribution of the vehicle delay tags is output by using the mixed density network; and obtaining the vehicle delay distribution based on the distribution of the vehicle delay labels. Compared with the prior art, the method has the advantages of high estimation precision, high traffic scene adaptability, low data quality requirement and the like.

Description

Technical Field

[0001] The invention relates to the field of traffic signal control and traffic parameter estimation, and in particular to a method for estimating vehicle delay distribution at a signal intersection. Background Art

[0002] Adaptive control of signalized intersections is considered to be the most basic means to improve traffic efficiency. Vehicle delay distribution is an important basis for evaluating and optimizing intersection signal control schemes. Connected vehicle trajectories provide detailed driving information of vehicles near intersections, and mining the required traffic information from them is a type of traffic technology with great application prospects. The information richness of connected vehicle trajectories also brings about the complexity of data structure, so the method of extracting information from them needs to be carefully designed. The Chinese patent application "CN116226592A" provides an intersection queue length estimation method based on sparse connected vehicle trajectory data, which estimates the intersection queue length by using sparse connected vehicle trajectory data, and constructs a speed matrix under the red light state, a speed matrix under the green light state, and a standard deviation matrix of the speed of the intersection section based on the connected vehicle trajectory, and estimates the intersection queue length based on the above-constructed matrix. Although it can effectively infer the queue length in the intersection, the calculation complexity involved in the process of constructing the matrix is ​​high, which is not conducive to real-time estimation. Although it considers many aspects of data, it ignores the impact of the penetration rate of connected vehicles on the queue length estimation.

[0003] Therefore, providing a method that can simplify the calculation process and take the penetration rate of connected vehicles into consideration is a technical problem that needs to be solved. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a method for estimating vehicle delay distribution at a signalized intersection. The method mines feature information based on a convolutional neural network and estimates vehicle delay distribution at a signalized intersection based on a mixed density network. The method has the advantages of high estimation accuracy and low data quality requirements.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for estimating vehicle delay distribution at a signalized intersection, wherein the method uses a vehicle delay distribution model to estimate vehicle delay distribution at a signalized intersection based on connected vehicle trajectory data, wherein the vehicle delay distribution model includes a convolutional neural network and a mixed density network, and the method includes:

[0007] Constructing a connected vehicle position matrix based on the connected vehicle trajectory data, and stacking the connected vehicle position matrix to obtain input features;

[0008] Inputting the input features into the convolutional neural network to extract feature data;

[0009] Based on the feature data, using the mixed density network to output the distribution of vehicle delay labels;

[0010] The vehicle delay distribution is obtained based on the distribution of the vehicle delay labels.

[0011] As a preferred technical solution, the method for constructing a networked vehicle position matrix includes:

[0012] Obtaining the time information of the networked vehicle trajectory data, and dividing all entrance roads of the intersection into grids based on the time information and the networked vehicle trajectory data;

[0013] According to the result of grid division, it is determined whether there is a vehicle in each grid. If so, the corresponding grid element is set to 1; if not, the corresponding grid element is set to 0, thereby generating the networked vehicle position matrix.

[0014] As a preferred technical solution, the stacking processing method includes:

[0015] Dividing the time information into a plurality of signal periods;

[0016] Obtain the position matrix of all connected vehicles in the same signal cycle and stack them in the time dimension to generate a three-dimensional tensor of time-lane-distance;

[0017] Obtain the three-dimensional tensors of all different signal periods, and perform time dimension splicing to generate the input features.

[0018] As a preferred technical solution, the method of using the mixed density network to output the distribution of vehicle delay labels is:

[0019] The mixed distribution component of the mixed density network is generated, and its expression is:

[0020]

[0021] Among them, z represents the vehicle delay label; x is the feature data; m k (x) is the expected vector of Gaussian component distribution; S k (x) is the diagonal covariance matrix of the Gaussian component distribution where s k.m is the standard deviation of the mth dimension of the kth Gaussian component distribution, where C is the dimension of the Gaussian component distribution;

[0022] The distribution of the vehicle delay label is estimated based on the mixed distribution component, and its expression is:

[0023]

[0024] Among them, p(z|x) is the distribution of vehicle delay labels under the condition of feature data x, K is the number of components of Gaussian mixture distribution, and p k (x) is the mixing coefficient of the kth Gaussian component distribution, p k (z|x) is the mixed distribution component of the kth mixed density network.

[0025] As a preferred technical solution, the training method of the vehicle delay distribution model is:

[0026] Acquire historical networked vehicle trajectory data, and acquire historical input features based on the historical networked vehicle trajectory data;

[0027] Calculating a vehicle delay value based on the historical networked vehicle trajectory data, and generating a vehicle delay label based on the vehicle delay value;

[0028] Iterate the following steps until the training stop condition is met:

[0029] Randomly select a batch of historical input features, input them into the vehicle delay distribution model to be trained, and output the corresponding historical vehicle delay probability density;

[0030] Calculating a loss value based on the historical vehicle delay probability density and the vehicle delay label;

[0031] Perform back-propagation to calculate the model parameter gradients;

[0032] The model parameters of the vehicle delay distribution model are updated based on the model parameter gradient optimization.

[0033] As a preferred technical solution, the method for generating a vehicle delay label is:

[0034] Obtaining historical time information of the historical networked vehicle trajectory data, and obtaining the time when the vehicle arrives at and leaves the intersection based on the historical time information;

[0035] Calculating vehicle travel time based on the times when the vehicle arrives at and leaves the intersection;

[0036] Obtain the expected speed and travel distance of the vehicle, and calculate the expected travel time of the vehicle;

[0037] Calculating a vehicle delay value based on the vehicle travel time and the vehicle expected travel time;

[0038] The vehicle delay value is transformed to generate the vehicle delay label.

[0039] As a preferred technical solution, the transformation includes power transformation and batch normalization, and its expression is:

[0040] z=s -1 (y l -m),

[0041] Among them, z represents the vehicle delay label, λ represents the parameter of the power transformation, s represents the standard deviation of the batch normalization transformation estimate, μ represents the expectation of the batch normalization transformation, and y represents the vehicle delay value.

[0042] As a preferred technical solution, the method for obtaining the vehicle delay distribution is: performing an inverse transformation on the distribution of the vehicle delay labels.

[0043] As a preferred technical solution, the training stop condition includes: the percentage of decrease of the loss function in this iteration compared with the loss function before this iteration is less than a preset value; or the number of iterations reaches the maximum number of iterations.

[0044] As a preferred technical solution, the calculation method of the loss value is:

[0045]

[0046] Where x' represents the historical input features, x represents the feature data, z represents the vehicle delay label, y represents the vehicle delay value, D represents the data set consisting of the historical input features and the vehicle delay labels, p(z|x) is the distribution of the vehicle delay labels under the condition of the feature data x, λ is the parameter of the power transformation, s is the standard deviation of the batch normalization transformation estimate, and μ is the expectation of the batch normalization transformation.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1) The present invention constructs a networked vehicle position matrix through grid division, stacks the position matrix periodically, obtains input features containing time and space information, maintains the rich information contained in the sparse networked vehicle trajectory data to the greatest extent, and can automatically extract the information required for vehicle delay distribution estimation from it. Compared with the existing traffic parameter estimation method based on networked vehicle trajectory, the data processing process is simpler and faster, and has higher estimation accuracy;

[0049] 2) Based on the convolutional neural network and mixed density network in the deep learning method, the present invention realizes the modeling of vehicle delay distribution in a data-driven manner, which can be applied to complex and changeable traffic conditions. Compared with the existing analytical estimation method of vehicle delay distribution, it has lower requirements on data quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2It is a flow chart of trajectory data processing of the present invention;

[0052] Figure 3 A flow chart for obtaining input features of the present invention;

[0053] Figure 4 This is a flow chart of the vehicle delay distribution model training of the present invention. DETAILED DESCRIPTION

[0054] 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 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 should fall within the scope of protection of the present invention.

[0055] Deep learning uses one or more neural networks to approximate the mapping relationship between traffic status and corresponding vehicle delay distribution. Through a large amount of data-driven learning, the neural network fits this mapping relationship so that the output estimation result is close to the real result. Because deep learning has the advantages of fast calculation and description of complex problems, the present invention proposes a signal intersection vehicle delay distribution estimation method based on a mixed density network. The process is as follows: Figure 1 As shown in the figure, based on the collected connected vehicle trajectory data, they are processed into input features and output labels that constitute the data set. Secondly, a vehicle delay distribution estimation model is constructed based on a convolutional neural network and a mixed density network. The result obtained by inputting the data into the model is used to calculate the negative log-likelihood loss function, and then the back-propagation process is performed to calculate the gradient of the model parameters. Finally, the gradient-based Adam algorithm is used to update the model parameters. The learning process is repeatedly iterated to guide convergence. The mixed density network is used to learn the mapping relationship between the traffic conditions reflected by the connected vehicle trajectory and the vehicle delay distribution at the signal intersection. By introducing improved mechanisms such as connected vehicle trajectory processing, delay transformation, and mixed density networks, sparse connected vehicle trajectory data can be effectively mined and the delay distribution of connected vehicles at signal intersections can be estimated, highlighting the advantages of strong versatility and high computational efficiency. The detailed steps are as follows:

[0056] S1. Data processing:

[0057] The present invention processes the acquired connected vehicle trajectory data in two parts: Figure 2 As shown, one is to obtain input features based on the connected vehicle trajectory data, and the other is to obtain vehicle delay labels based on the connected vehicle trajectory data.

[0058] S11. Obtain input features:

[0059] S111, obtaining the time information of the networked vehicle trajectory data, and dividing all entrance roads of the intersection into grids based on the time information and the networked vehicle trajectory data, such as Figure 3 shown.

[0060] S112. According to the result of grid division, determine whether there is a vehicle in each grid. If so, set the corresponding grid element to 1; if not, set the corresponding grid element to 0, and generate a networked vehicle position matrix.

[0061] S113. Divide the time information into multiple signal periods.

[0062] S114. Obtain the position matrices of all connected vehicles in the same signal cycle, and stack them in the time dimension to generate a three-dimensional tensor of time-lane-distance.

[0063] S115. Obtain three-dimensional tensors of all different signal periods, and perform time dimension splicing to generate input features.

[0064] S12. Obtain vehicle delay label:

[0065] S121. Obtain historical time information of historical networked vehicle trajectory data, and obtain the time when the vehicle arrives at and leaves the intersection based on the historical time information.

[0066] S122. Calculate the vehicle travel time based on the time when the vehicle arrives at and leaves the intersection.

[0067] S123: Obtain the expected speed and travel distance of the vehicle, and calculate the expected travel time of the vehicle.

[0068] S124. Subtract the vehicle travel time from the vehicle expected travel time to obtain a vehicle delay value.

[0069] S125. Perform power transformation and batch normalization on the vehicle delay value to generate a vehicle delay label, which is expressed as:

[0070] z=s -1 (y l -m),

[0071] Among them, z represents the vehicle delay label, λ represents the parameter of the power transformation, s represents the standard deviation of the batch normalization transformation estimate, μ represents the expectation of the batch normalization transformation, and y represents the vehicle delay value.

[0072] S2. Input the input features into the convolutional neural network to extract feature data, and use the feature data as the input of the mixed density network.

[0073] S3. Use the mixed density network to output the distribution of vehicle delay labels:

[0074] S31. Generate the mixed distribution component of the mixed density network, which is expressed as:

[0075]

[0076] Among them, z represents the vehicle delay label; x is the feature data; m k (x) is the expected vector of Gaussian component distribution; S k (x) is the diagonal covariance matrix of the Gaussian component distribution where s k.m is the standard deviation of the mth dimension of the kth Gaussian component distribution, where C is the dimension of the Gaussian component distribution;

[0077] S32. Estimate the distribution of vehicle delay labels based on mixed distribution components, and its expression is:

[0078]

[0079] Among them, p(z|x) is the distribution of vehicle delay labels under the condition of feature data x, K is the number of components of Gaussian mixture distribution, and p k (x) is the mixing coefficient of the kth Gaussian component distribution, p k (z|x) is the mixed distribution component of the kth mixed density network.

[0080] S4. Perform an inverse transformation on the distribution of vehicle delay labels to obtain vehicle delay distribution.

[0081] In the present invention, the training process of the vehicle delay distribution model is as follows: Figure 4 As shown, the following steps are included:

[0082] A1. Obtain historical connected vehicle trajectory data, and obtain historical input features based on the historical connected vehicle trajectory data.

[0083] A2. Construct a vehicle delay label according to the process of step S12.

[0084] Iterate steps A3-A6 until the percentage of the loss function in this iteration compared to the loss function before this iteration is less than a preset value, or the number of iterations reaches the maximum number of iterations:

[0085] A3. Randomly select a batch of historical input features, input them into the vehicle delay distribution model to be trained, and output the corresponding historical vehicle delay probability density.

[0086] A4. Calculate the loss value based on the historical vehicle delay probability density and vehicle delay labels. The expression is:

[0087]

[0088] Where x' represents the historical input features, x represents the feature data, z represents the vehicle delay label, y represents the vehicle delay value, D represents the data set consisting of the historical input features and the vehicle delay labels, p(z|x) is the distribution of the vehicle delay labels under the condition of the feature data x, λ is the parameter of the power transformation, s is the standard deviation of the batch normalization transformation estimate, and μ is the expectation of the batch normalization transformation.

[0089] A5. Perform back propagation to calculate the model parameter gradients.

[0090] A6. Based on the model parameter gradient, the Adam algorithm is used to optimize and update the model parameters of the vehicle delay distribution model. The specific parameter update form is:

[0091]

[0092] Where θ is the model parameter; L is the loss value; α is the learning rate.

[0093] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for estimating vehicle delay distribution at a signalized intersection, characterized in that: The method uses a vehicle delay distribution model to estimate vehicle delay distribution at a signalized intersection based on connected vehicle trajectory data, wherein the vehicle delay distribution model includes a convolutional neural network and a mixed density network, and the method includes: Constructing a connected vehicle position matrix based on the connected vehicle trajectory data, and stacking the connected vehicle position matrix to obtain input features; Inputting the input features into the convolutional neural network to extract feature data; Based on the feature data, using the mixed density network to output the distribution of vehicle delay labels; The vehicle delay distribution is obtained based on the distribution of the vehicle delay labels.

2. A method for estimating vehicle delay distribution at a signalized intersection according to claim 1, characterized in that: The method for constructing a networked vehicle position matrix includes: Obtaining the time information of the networked vehicle trajectory data, and dividing all entrance roads of the intersection into grids based on the time information and the networked vehicle trajectory data; According to the result of grid division, it is determined whether there is a vehicle in each grid. If so, the corresponding grid element is set to 1; if not, the corresponding grid element is set to 0, thereby generating the networked vehicle position matrix.

3. A method for estimating vehicle delay distribution at a signalized intersection according to claim 2, characterized in that: The stacking method includes: Dividing the time information into a plurality of signal periods; Obtain the position matrix of all connected vehicles in the same signal cycle and stack them in the time dimension to generate a three-dimensional tensor of time-lane-distance; Obtain the three-dimensional tensors of all different signal periods, and perform time dimension splicing to generate the input features.

4. The method for estimating vehicle delay distribution at a signalized intersection according to claim 1, characterized in that: The method of using the mixed density network to output the distribution of vehicle delay labels is: The mixed distribution component of the mixed density network is generated, and its expression is: Among them, z represents the vehicle delay label; x is the feature data; m k (x) is the expected vector of Gaussian component distribution; S k (x) is the diagonal covariance matrix of the Gaussian component distribution where s k.m is the standard deviation of the mth dimension of the kth Gaussian component distribution, where C is the dimension of the Gaussian component distribution; The distribution of the vehicle delay label is estimated based on the mixed distribution component, and its expression is: Among them, p(z|x) is the distribution of vehicle delay labels under the condition of feature data x, K is the number of components of Gaussian mixture distribution, and p k (x) is the mixing coefficient of the kth Gaussian component distribution, p k (z|x) is the mixed distribution component of the kth mixed density network.

5. The method for estimating vehicle delay distribution at a signalized intersection according to claim 1, characterized in that: The training method of the vehicle delay distribution model is: Acquire historical networked vehicle trajectory data, and acquire historical input features based on the historical networked vehicle trajectory data; Calculating a vehicle delay value based on the historical networked vehicle trajectory data, and generating a vehicle delay label based on the vehicle delay value; Iterate the following steps until the training stop condition is met: Randomly select a batch of historical input features, input them into the vehicle delay distribution model to be trained, and output the corresponding historical vehicle delay probability density; Calculating a loss value based on the historical vehicle delay probability density and the vehicle delay label; Perform back-propagation to calculate the model parameter gradients; The model parameters of the vehicle delay distribution model are updated based on the model parameter gradient optimization.

6. A method for estimating vehicle delay distribution at a signalized intersection according to claim 5, characterized in that: The method for generating a vehicle delay label is: Obtaining historical time information of the historical networked vehicle trajectory data, and obtaining the time when the vehicle arrives at and leaves the intersection based on the historical time information; Calculating vehicle travel time based on the times when the vehicle arrives at and leaves the intersection; Obtain the expected speed and travel distance of the vehicle, and calculate the expected travel time of the vehicle; Calculating a vehicle delay value based on the vehicle travel time and the vehicle expected travel time; The vehicle delay value is transformed to generate the vehicle delay label.

7. A method for estimating vehicle delay distribution at a signalized intersection according to claim 6, characterized in that: The transformation includes power transformation and batch normalization, and its expression is: z=s -1 (y l -m), Among them, z represents the vehicle delay label, λ represents the parameter of the power transformation, s represents the standard deviation of the batch normalization transformation estimate, μ represents the expectation of the batch normalization transformation, and y represents the vehicle delay value.

8. A method for estimating vehicle delay distribution at a signalized intersection according to claim 7, characterized in that: The method for obtaining the vehicle delay distribution is: performing an inverse transformation on the distribution of the vehicle delay labels.

9. The method for estimating vehicle delay distribution at a signalized intersection according to claim 5, characterized in that: The training stop conditions include: the percentage of the loss function in this iteration compared with the loss function before this iteration is less than a preset value; or the number of iterations reaches the maximum number of iterations.

10. The method for estimating vehicle delay distribution at a signalized intersection according to claim 5, characterized in that: The calculation method of the loss value is: Where x' represents the historical input features, x represents the feature data, z represents the vehicle delay label, y represents the vehicle delay value, D represents the data set consisting of the historical input features and the vehicle delay labels, p(z|x) is the distribution of the vehicle delay labels under the condition of the feature data x, λ is the parameter of the power transformation, s is the standard deviation of the batch normalization transformation estimate, and μ is the expectation of the batch normalization transformation.

Images