Vehicle following behavior prediction method based on deep learning and multi-dimensional fusion data

Abstract

The present invention discloses a method for predicting vehicle following behavior based on deep learning and multi-dimensional fusion data, comprising the following steps: extracting unstructured data of other vehicles in the driver's perception area based on microscopic traffic flow trajectory data; encoding the unstructured data using the DIM model based on information entropy theory; fusing the encoded unstructured data with the structured data; inputting the fused data into a deep learning model to perform microscopic following behavior prediction and output the travel distance of the following vehicle in the next time period. The present invention proposes a method for predicting following behavior that simultaneously considers both unstructured and structured features, and the prediction method has excellent performance in both single-step and multi-step prediction tasks; the method of the present invention has strong theoretical and operational properties, and by fusing unstructured data with structured data, the vehicle's perception of the surrounding environment becomes more comprehensive and accurate.

Description

Technical Field

[0001] The present invention relates to the field of data-driven microscopic traffic flow, and in particular to a vehicle following behavior prediction method based on deep learning and multi-dimensional fusion data. Background Art

[0002] Emerging technologies represented by big data have become a new driving force for the development of my country's smart transportation industry. Against this background, the abundance of traffic big data resources and the advancement of traffic big data analysis and mining technologies have had a profound impact on traffic flow theory and methods. It is necessary to reshape the traffic flow theory system under the big data environment and realize the detailed modeling of traffic flow.

[0003] As an essential component of traffic flow research, car-following models provide theoretical support for both microscopic and macroscopic traffic flow simulation models. With advances in data acquisition technology and improved data processing capabilities, data-driven car-following models have gradually emerged. Data-driven approaches do not require assumptions about microscopic traffic behavior, are more flexible, and can incorporate a wider range of information.

[0004] Existing data-driven car-following models have achieved good results in theoretical research, but most still use the three parameters of relative vehicle speed, speed, and headway in traditional models to analyze car-following behavior. These models do not expand on the dimensions of the vehicle's microscopic level, fail to fully utilize the advantages of data-driven models, and are likely unable to accurately predict the vehicle's speed at the next moment. Summary of the Invention

[0005] Purpose of the invention: In response to the above problems, the purpose of the present invention is to provide a vehicle following behavior prediction method based on deep learning and multi-dimensional fusion data.

[0006] Technical solution: The present invention provides a method for predicting vehicle following behavior based on deep learning and multi-dimensional fusion data, the method comprising the following steps:

[0007] Step 1: Extracting unstructured data of other vehicles in the driver's perception area based on microscopic traffic flow trajectory data; the unstructured data includes location, speed, number, and size;

[0008] Step 2: Encode unstructured data using the DIM model based on information entropy theory;

[0009] Step 3: Fusing the encoded unstructured data and structured data; the fusion includes splicing the two based on the last axis;

[0010] Step 4: Input the spliced ​​data into the deep learning model to predict the micro-following behavior and output the driving distance of the following vehicle in the next period.

[0011] Furthermore, step 2 specifically includes:

[0012] The DIM model is trained using unstructured data from other vehicles. The encoder in the DIM model takes maximum mutual information and statistical constraints as its goals, and uses the back-propagation algorithm to iteratively optimize the DIM model parameters, find the optimal encoder in the encoder set, and output the encoding of the unstructured data.

[0013] Furthermore, the deep learning model is a DIMNN model, which includes 3 fully connected layers, and each fully connected layer also includes batch normalization and Relu activation function.

[0014] Furthermore, the structured data includes the speed of the following vehicle, the speed difference with the preceding vehicle, the headway between the following vehicle and the preceding vehicle, and the speed change and position change of the preceding vehicle.

[0015] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention proposes a method for predicting car-following behavior that takes into account both unstructured and structured features. Based on the concept of information entropy, the unstructured data is encoded, and three objectives, namely global mutual information, local mutual information and prior distribution, are established to optimize model parameters. Compared with traditional models, the prediction method has excellent performance in both single-step prediction and multi-step prediction tasks; the method of the present invention has strong theoretical and operational properties, and by integrating unstructured data with structured data, the vehicle's perception of the surrounding environment is more comprehensive and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the vehicle following behavior prediction method of the present invention;

[0017] Figure 2 It is a schematic diagram of four sets of unstructured data;

[0018] Figure 3 This is a demonstration of the multi-step prediction effect of the car-following model considering unstructured features. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of this application more clear, this application is further described in detail below with reference to the accompanying drawings and embodiments.

[0020] like Figure 1 The vehicle following behavior prediction method based on deep learning and multi-dimensional fusion data described in this embodiment includes the following steps:

[0021] Step 1: Extract unstructured data of other vehicles in the driver's perception area based on microscopic traffic flow trajectory data.

[0022] The state of a car following is not only related to the current vehicle's state and the state of the vehicle ahead, but also to the road traffic conditions perceived by the current vehicle. To better capture the factors influencing car following behavior, unstructured data is introduced to graphically depict the current vehicle's perception of road traffic conditions at a given moment. This unstructured data includes the location, speed, number, and size of surrounding vehicles.

[0023] This example uses high-precision Next Generation Simulation (NGSIM) data from real-world vehicle trajectories for algorithm construction and case analysis. The NGSIM dataset is collected by a fixed-position camera and then extracted through video processing. The sampling frequency is 0.1 seconds, and each sample includes information such as the vehicle's instantaneous velocity, acceleration, longitudinal and lateral position coordinates, vehicle length, and vehicle type. The driving data for vehicles numbered 684, 744, and 2695 were selected. Based on the typical driver's perceived range while driving, the perception area was divided into 0.2m x 0.2m grids, 20m in front of and behind the driver and 8m to the left and right. This results in an unstructured data size of 200 x 80 units. The grids are filled with the vehicle's speed if the corresponding grid location is occupied; 0 if the corresponding grid location is on the road; and -1 if the corresponding grid location is not drivable (e.g., the shoulder). Figure 2 Figure 1 shows four sets of unstructured data. Vehicles are represented by rectangles, with the following vehicle in question placed in the center. In each of the four data sets, the number of vehicles within the driver's perception range for the following vehicle is 6, 3, 5, and 6, respectively. The perception ranges of the first and third vehicles include unpassable areas, and the vehicles around the first, second, and fourth vehicles are either faster than or similar in speed.

[0024] Before encoding, invalid data in unstructured data is filtered out, such as grids filled with -1, to speed up data reprocessing.

[0025] Step 2: Use the DIM model to encode unstructured data based on information entropy theory.

[0026] The DIM (Deep InfoMax) model is trained using unstructured data from other vehicles as training sample data. The encoder in the DIM model takes maximizing mutual information and statistical constraints as its goals, and uses the back-propagation algorithm to iteratively optimize the DIM model parameters, find the optimal encoder in the encoder set, and output the encoding of the unstructured data.

[0027] The DIM model consists of a feature mapper, an encoder, and three discriminators, namely global mutual information, local mutual information, and prior distribution. The specific parameters of each part are shown in Table 1.

[0028] Through the encoder set ε Φ To encode the unstructured data in step 1, the encoder set is represented as ε Φ ={E ψ} ψ∈Ψ , where E ψ : Indicates from space To Space The mapping parameter is ψ; training samples That is, unstructured data, and the training samples obey the empirical probability distribution P. Define U ψ，P For samples obtained from distribution P through encoder E ψ The subsequent distribution, i.e., encoding The goal of model training is to get the distribution of Φ Find the best E ψ , so that the encoding Better representation of unstructured data with fewer data dimensions. The encoder is trained according to the following objectives:

[0029] (1) Maximum mutual information: Find a set of parameters ψ that maximizes the global mutual information and local mutual information, ensuring that the encoder output better represents the unstructured data sample;

[0030] (2) Statistical constraints: U ψ，P should be consistent with the prior distribution The output of the encoder has certain prior features.

[0031] Table 1 DIM model structure

[0032]

[0033]

[0034] Step 3: Fuse the encoded unstructured data with the structured data.

[0035] The unstructured data encoded by other vehicles and the microscopic structured data of the current vehicle are spliced ​​based on the last axis. The encoding tensor size of the unstructured data is 32×1, the tensor size of the structured data is 5×1, and the tensor size of the spliced ​​fused data is 37×1. The fused data is expressed in the form of [E ψ (x i (t)),v i (t),s i (t),Δv i (t),Δx i-1 (t-1),dv i-1 (t)], where Eψ (x i (t)) represents the encoded unstructured data, v i (t) represents the speed of vehicle i at time t, s i (t) represents the headway between vehicle i at time t, Δv i (t) = v i (t)-v i-1 (t) represents the speed difference between vehicle i and the preceding vehicle at time t, Δx i-1 (t-1) is the change in the position of the preceding vehicle, indicating the distance the preceding vehicle has traveled from time t-1 to time t, dv i-1 (t) = v i-1 (t)-v i-1 (t-1) is the speed change of the preceding vehicle, which indicates the speed difference between the preceding vehicle at time t and time t-1.

[0036] The structured data includes the speed of the following vehicle, the speed difference and headway between the following vehicle and the preceding vehicle, as well as the speed change and position change of the preceding vehicle. The first three variables of the structured data are classic variables in the following model, while the last two variables represent the following vehicle's perception of the changes in the preceding vehicle's behavior.

[0037] Step 4: Input the spliced ​​data into the deep learning model to predict the micro-following behavior and output the driving distance of the following vehicle in the next period.

[0038] The deep learning model used is a DIMNN model, consisting of three fully connected layers with input and output dimensions of (37, 32), (32, 16), and (16, 1), respectively. Each fully connected layer also includes batch normalization and a Relu activation function. The final output is a one-dimensional prediction of the next distance traveled by the following vehicle.

[0039] To further validate the effectiveness and accuracy of the vehicle following behavior prediction method of the present invention, this embodiment uses a gated recurrent unit (GRU) model and an intelligent driver model (IDM) as benchmark models, performing single-step and multi-step predictions, respectively. Single-step prediction predicts the following behavior at the next time step, given the current state of the vehicle under study and other necessary parameters. Each time step prediction is based on real data, and the mean absolute error between the current vehicle's speed and position is used as the core metric. Multi-step prediction predicts the following behavior of the vehicle multiple time steps into the future, given the initial state of the vehicle under study and other necessary parameters. Each time step prediction is based on the prediction results of the previous time step. The following vehicle's operating state is then iteratively updated, along with the actual information of the preceding vehicle and the position and speed information of surrounding vehicles, to sequentially predict the following behavior of the subject multiple time steps into the future. Multi-step prediction uses the Euclidean distance (EUC), Fréchet distance (FRE), area error ratio (QE), and the farthest moment position prediction accuracy (FD) between the actual and predicted vehicle trajectories as metrics.

[0040] The vehicle following behavior prediction method based on deep learning and multi-dimensional fusion data described in the present invention exceeds the baseline model in speed prediction accuracy in single-step prediction tasks, and is comparable to the IDM model in position prediction accuracy. In multi-step prediction tasks, the model is significantly better than the baseline model. Compared with the best performing baseline models in various indicators, the mean value of the EUC indicator is only about 70% of the GRU model, the mean value of the FRE indicator is only about 65% of the IDM model, the mean value of the QE indicator is only about 70% of the GRU model, and the mean value of the FD indicator is only about 44% of the IDM model. Considering the stability of the model performance, the upper quartiles of the speed and position MAE of the single-step prediction task are comparable to those of the IDM model, the upper quartile of the EUC of the multi-step prediction task is slightly better than the GRU model, the upper quartiles of the FRE indicator and the FD indicator are far better than all baseline models, and the upper quartile of the QE indicator is slightly better than the GRU model. The vehicle driving data with vehicle numbers 684, 744 and 2695 are selected, and their multi-step prediction results are visualized, as shown below. Figure 3, the solid lines respectively represent the changes in the position or speed of the leading vehicle and the position or speed of the following vehicle (research object) over time, and the dotted line represents the predicted changes in the position or speed of the following vehicle (research object) over time. According to the different driving times of the following vehicles on the road section, the time span of the multi-step prediction is set to about 80s. Combined with the three trajectory data, the model can accurately predict the vehicle operation status and speed changes. The following prediction method considering unstructured data proposed in the present invention performs more prominently in the stability of single-step prediction, and is significantly better than the baseline model in the multi-step prediction task. The excellent performance of the prediction model in the present invention is mainly due to the introduction of unstructured data, and the vehicle's perception of the environment is more comprehensive and accurate.

[0041] Table 2 Prediction structure evaluation

[0042]

Claims

1. A vehicle following behavior prediction method based on deep learning and multi-dimensional fusion data, characterized by: The method comprises the following steps: Step 1: Extracting unstructured data of other vehicles in the driver's perception area based on microscopic traffic flow trajectory data; the unstructured data includes location, speed, number, and size; Step 2: Encode unstructured data using the DIM model based on information entropy theory; Step 3: Fusing the encoded unstructured data and the structured data; the fusion includes splicing the two based on the last axis; Step 4: Input the spliced ​​data into the deep learning model to predict microscopic car-following behavior and output the distance traveled by the following vehicle in the next period. Step 2 specifically includes: The DIM model is trained using unstructured data from other vehicles. The encoder in the DIM model takes maximum mutual information and statistical constraints as its goals, and uses the back-propagation algorithm to iteratively optimize the DIM model parameters, find the optimal encoder in the encoder set, and output the encoding of the unstructured data.

2. The vehicle following behavior prediction method according to claim 1, characterized in that: The deep learning model is a DIMNN model, which includes three fully connected layers, and each fully connected layer also includes batch normalization and Relu activation function.

3. The vehicle following behavior prediction method according to claim 1, characterized in that: The structured data includes the speed of the following vehicle, the speed difference with the preceding vehicle, the headway between the following vehicle and the preceding vehicle, and the speed change and position change of the preceding vehicle.

Images