Multi-dimensional cooperative vehicle lane changing decision-making method based on user portraits

Abstract

The invention relates to a multi-dimensional cooperative vehicle lane changing decision-making method based on a user portrait, and belongs to the technical field of mobile communication. In an automatic driving technology, a traditional method of a neural network in vehicle lane changing decision mainly depends on single vehicle data, so that the accuracy and the real-time performance are poor. In order to solve the problem, firstly, a vehicle generates a driving characteristic graph and a vehicle trajectory flow graph based on sensor data, and shares the driving characteristic graph and the vehicle trajectory flow graph to surrounding vehicles and a road side unit to realize data collaboration; secondly, the vehicle downloads the initial decision model from the road side unit, and carries out lane changing decision by combining data of the vehicle and surrounding vehicles; then, based on the data and traffic factors, the road side unit uses a deep neural network to update a model and re-issues the model to the vehicle, thereby ensuring real-time performance and safety of decision making. According to the method, the user portrait is generated by integrating the historical driving characteristics and the vehicle trajectory prediction, and then the deep neural network is applied, so that more accurate lane changing decision is realized, and the safety and the operation efficiency of a traffic system are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile communications and relates to a multi-dimensional collaborative vehicle lane change decision method based on user portraits. Background Art

[0002] With the development of vehicle networking and intelligent driving technology, information sharing and collaborative decision-making have become crucial in improving traffic safety and efficiency. The existing lane change decision-making system mainly relies on single-vehicle data, which has limitations in providing environmental perception and is difficult to fully capture the surrounding complex traffic conditions, thus restricting the accuracy and safety of decision-making results.

[0003] Traditional lane change decision systems usually rely on the vehicle's own sensor data, such as speed, driving trajectory, steering angle, etc. However, these data have limitations and can only provide the vehicle's own movement information, but cannot fully understand other vehicles and dynamic changes in the surrounding environment. Especially on busy urban roads or highways, the perspective of a single vehicle is difficult to ensure the comprehensiveness and safety of lane change decisions.

[0004] In the Internet of Vehicles environment, information sharing between vehicles is the key to solving this problem. By sharing real-time information such as location, speed, and driving trajectory with surrounding vehicles, vehicles can obtain more comprehensive traffic data and make more optimized lane change decisions. This information sharing requires efficient, low-latency communication technology, as well as reliable data processing and decision-making models.

[0005] Deep neural networks are excellent at processing complex data patterns and performing efficient learning, and have been widely used in the field of autonomous driving. However, existing decision-making models often lack the ability to update in real time during lane change decisions, making it difficult to adapt to rapidly changing traffic environments. In particular, when traffic conditions change suddenly, existing models cannot be adjusted in time, which may lead to safety hazards.

[0006] In addition, the user's driving style also has an important impact on lane change decisions. Driving style includes the driver's behavioral characteristics such as acceleration, braking, and steering, which can be described by parameters such as the vehicle's speed, acceleration, and steering wheel angle. The driving styles of different drivers will lead to significant differences in the decision-making behavior of vehicles under similar traffic conditions. For example, some drivers may prefer to drive conservatively and choose to change lanes less frequently, while other drivers may be more aggressive and change lanes frequently in order to seek faster driving speeds. The differences in user profiles are not only reflected between individual drivers, but are also affected by different traffic conditions, road conditions, and driving purposes. In order to improve the personalization and accuracy of lane change decisions, user profiles need to be incorporated into the decision model.

[0007] The present invention aims to solve the above problems and proposes a prediction-collaboration-decision-making integrated vehicle lane change decision method based on user portrait. First, the user driving behavior characteristics in the historical data are analyzed, and a personalized driving style model is constructed to perform user portraits. According to different user portraits, the system makes lane change decisions that are more in line with the driver's expectations, thereby improving user experience and safety. Secondly, through vehicle-road collaboration, the vehicle's local data is shared with the real-time information of surrounding vehicles, and decisions are adjusted based on the traffic flow diagram predicted by the vehicle trajectory, so that the CAV can obtain a more comprehensive perception of the traffic environment, quickly adapt to changes in traffic conditions, and reduce potential safety risks. Finally, deep learning models (such as ResNet and Transformer) are used to process and extract complex environmental features, learn key patterns and features from large amounts of traffic data, and improve the accuracy of lane change decisions. Summary of the invention

[0008] In view of this, the purpose of the present invention is to provide a multi-dimensional collaborative vehicle lane change decision method based on user portraits. This method deeply integrates the historical driving data of connected and autonomous vehicles (CAV) and surrounding vehicles (Surrounding vehicles, SVs), constructs a driving feature picture (DFP) and a trajectory flow diagram (TFD), and uses deep neural networks to provide more accurate and safe decision support for vehicle lane changes through data sharing of road side units (RSU), aiming to achieve more accurate and efficient lane change decisions through historical driving characteristics, vehicle trajectory prediction, and vehicle data sharing. This method combines the vehicle's local sensor data, the edge model trained by the RSU, and the shared information of surrounding vehicles to form a closed-loop system of prediction-cooperation-decision-making, thereby improving the safety and operating efficiency of the overall transportation system.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] In a first aspect, an embodiment of the present invention provides a multi-dimensional collaborative vehicle lane change decision method based on user portraits according to the characteristics of a vehicle lane change scenario. The method comprises the following steps:

[0011] S1: Data sharing model based on CAV-SVs-RSU

[0012] S2: CAV DFP generation scheme based on user portrait;

[0013] S3: Lane change decision scheme based on DFP and TFD of CAVs and SVs;

[0014] S4: TFD generation scheme for CAV based on lane-changing maneuvers;

[0015] S5: DFP and TFD sharing scheme for CAV;

[0016] S6: Transformer network model based on ResNet.

[0017] Secondly, in S1, a data sharing model based on CAV-SVs-RSU is adopted to improve the safety of lane change decisions through data sharing between CAV, SVs and RSU. CAV and SVs share DFP and TFD to improve CAV's perception of the surrounding environment; CAV and RSU share data to improve the efficiency of model updating; when the accuracy of the new model is improved, CAV and RSU share edge models to improve the accuracy of lane change decisions.

[0018] On the third aspect, in S1 of the embodiment of the present invention, CAV calculates the lane change incentive parameter by comprehensively considering multiple factors such as the current lane achievable speed, the current vehicle speed, the distance to the preceding vehicle, the preceding vehicle time distance, etc. When the parameter reaches the preset threshold, the system executes the lane change decision process, and CAV generates a DFP representing the driving characteristics based on the sensor data. Specifically, CAV collects L feature data of its own sensors, and calculates its M statistical functions respectively, generates a two-dimensional matrix according to the calculation results, and generates a DFP reflecting the spatiotemporal historical driving characteristics of the vehicle through a sliding time window of length S seconds. In this way, specific features can be extracted from historical driving information to implicitly describe the user's driving style, and the data used for decision-making will also contain more specific driving style information, so that the user's driving style can be simulated more accurately.

[0019] Fourthly, in S2 of the embodiment of the present invention, the vehicle downloads the edge model from the RSU, receives the DFP and TFD of the SVs, and applies the edge decision model to make a lane change decision based on these data. The decision model is trained and updated by a deep neural network, and the SVs include the K vehicles closest to the CAVs, including the leading vehicle in the current lane, the leading vehicle in the left adjacent lane, the leading vehicle in the right adjacent lane, the following vehicle in the left adjacent lane, the following vehicle in the right adjacent lane, the parallel vehicles in the left adjacent lane, and the parallel vehicles in the right adjacent lane. The method for SVs to generate TFD and DFP is the same as that of CAV. The DFP and TFD of SVs can reflect the historical driving characteristics of SVs and the prediction of their future trajectory, thereby making the lane change decision of CAV safer and more accurate.

[0020] In the fifth aspect, in S3 of the embodiment of the present invention, the CAV generates a local TFD according to the output action of the lane change decision. The TFD is a two-dimensional grid matrix generated according to the lane change behavior of the vehicle. By encoding the spatiotemporal information of the grid matrix, the occupancy of obstacles and the change of the vehicle position over time (i.e., the future driving trajectory of the vehicle) are represented. Specifically, the grid positions occupied by obstacles are uniformly set to 1, the value of the vehicle position at time t is encoded as 1, and decreases linearly to 0 with time (i.e., at time t+T D The value of the position is 0, where T D is the time length of the trajectory). This provides a representation of the vehicle's demand for road rights. This encoding allows CAVs to make more fine-grained reasoning decisions about the driving behavior of SVs.

[0021] In the sixth aspect, in S4 of the embodiment of the present invention, the DFP and TFD of CAV and SVs are uploaded to the RSU experience playback pool for RSU edge model update. At the same time, the DFP and TFD of CAV will be shared with SVs for lane change decision of SVs. CAV promotes cooperation between vehicles and improves the safety of lane change decision model by sharing driving intention and lane change information with SVs.

[0022] In the seventh aspect, in embodiment S5 of the present invention, the RSU uses the DFP and TFD sampled from the experience replay pool, combined with the ResNet network of the attention mechanism to extract the historical driving feature information of the CAV, and converts it into a vector, which is connected with the 10-dimensional risk assessment vector obtained according to traffic factors in the RSU, and input into the decision network. The model accuracy is calculated according to the actual lane change decision action, and the model is updated when the edge model accuracy improves. The vehicle downloads the updated edge model from the RSU for local lane change decision-making, and uploads the new data generated by the decision to the RSU experience replay pool.

[0023] The beneficial effects of the present invention are:

[0024] (1) Improving the accuracy of lane change decisions:

[0025] By analyzing historical driving data and building a personalized driving style model, we can more accurately simulate the driving behavior of real users, make lane change decisions more in line with the driver's expectations and habits, and improve the user experience.

[0026] The fusion of local vehicle sensor data, edge models trained by roadside units, and shared information from surrounding vehicles provides the vehicle with a more comprehensive environmental perception, reducing the risk of misjudgment and collision.

[0027] Deep learning models are used to process and extract complex environmental features, learn key patterns and features from large amounts of traffic data, and improve the accuracy of lane change decisions.

[0028] (2) Enhanced lane change decision safety:

[0029] Through vehicle-road collaboration, vehicles share real-time data, including driving characteristic maps, trajectory flow maps, etc., to achieve information sharing and collaborative decision-making, enhance perception of the surrounding environment, and reduce safety hazards.

[0030] The roadside unit uses a deep learning model to continuously update the edge model based on the data in the experience playback pool and sends it to the vehicle to ensure the real-time and adaptability of the decision-making model.

[0031] The vehicle generates a trajectory flow graph based on the lane change decision, predicts the future driving trajectory, and characterizes the road right demand, helping the vehicle to more accurately judge the driving intentions of surrounding vehicles and avoid potential conflicts.

[0032] (3) Improving the efficiency of the transportation system:

[0033] Through more accurate and safer lane change decisions, traffic congestion can be reduced, traffic flow can be optimized and road traffic efficiency can be improved.

[0034] Reduce traffic accidents caused by improper lane changing operations, improve the safety of the traffic system, and reduce social costs.

[0035] Optimizing traffic flow can reduce vehicle fuel consumption and emissions, promote energy conservation and emission reduction, and protect the environment.

[0036] (4) Good scalability:

[0037] More types of data sources can be connected as needed, such as meteorological information, traffic event information, etc., to further improve the accuracy and adaptability of the decision-making model.

[0038] The model architecture and parameters can be adjusted according to different application scenarios and requirements to make it more suitable for different traffic environments and driving styles.

[0039] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0041] Figure 1 This is a system architecture diagram of a multi-dimensional collaborative vehicle lane change decision method based on user portrait;

[0042] Figure 2 Generate a flow chart for DFP of CAV based on historical driving data;

[0043] Figure 3 This is a schematic diagram of the Transformer network based on ResNet;

[0044] Figure 4 Execution flowchart of the multi-dimensional collaborative vehicle lane change decision method based on user portrait. DETAILED DESCRIPTION

[0045] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0046] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0047] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0048] Figure 1 The schematic diagram shows the whole process of vehicle data collection, data processing, model download, lane change decision, generation and sharing of traffic snapshots, data upload, risk assessment and model update. Figure 1As shown in the figure, the network considers the historical driving data of vehicle driving characteristics and the trajectory flow diagram of vehicle trajectory prediction. Vehicles share this real-time information to improve the lane change safety of CAV based on the decision model. At the same time, this information is uploaded to the RSU experience playback pool as new data to update the lane change model, and three traffic factors, safety, incentive factors, and tolerance factors, are considered.

[0049] 1. CAV DFP generation solution based on historical driving data

[0050] This method generates a two-dimensional matrix by calculating M statistical functions of vehicle driving characteristics, and uses time window extraction to convert the data into DFP. Therefore, specific features can be extracted from historical driving information to implicitly describe the user profile of driving characteristics, with the purpose of performing lane change decision tasks and including more specific driving style information. Figure 2 As shown in the figure, the specific process of CAV converting historical data into DFP is as follows:

[0051] (1) Data collection: The current vehicle j collects the relative position y through sensors. j , relative position x j , lateral speed Longitudinal speed Lateral acceleration Longitudinal acceleration Headway d jk 、Headway h jk etc. L historical driving features. Among them, k∈{B,P,PL,FL,FR,ASL,ASR}. B, P, PL, FL, FR, ASL and ASR represent the following vehicle in the current lane, the preceding vehicle in the current lane, the preceding vehicle in the left adjacent lane, the preceding vehicle in the right adjacent lane, the following vehicle in the left adjacent lane, the following vehicle in the right adjacent lane, the parallel vehicles in the left adjacent lane and the parallel vehicles in the right adjacent lane, respectively.

[0052] (2) Statistical function calculation: For the L driving characteristics, M statistical functions such as the mean value, standard deviation, median, 25% percentile, 75% percentile, minimum value, and maximum value are calculated for each second, and a column of a two-dimensional matrix is ​​generated based on the obtained results, that is, a column vector of size L·M×1 (L driving characteristics×M statistical functions) is generated every second.

[0053] (3) Time window setting: As time accumulates, the two-dimensional matrix gradually becomes larger, reflecting the long-term driving characteristics of the CAV. Since the lane change time is very short, the focus is mainly on short-term driving characteristics. Therefore, a time window of length S seconds is set to frame the two-dimensional matrix S seconds before the lane change, which means that the size of each DFP is L·M×S.

[0054] (4) Normalization: All data in the two-dimensional matrix are normalized to eliminate the impact of different feature dimensions on DFP.

[0055] (5) DFP generation: All data in the two-dimensional matrix are represented as different colors. The depth of the color reflects the size of the data, and the type of color represents the category of the driving feature, thereby generating a driving feature map and implicitly representing the driving feature as an image.

[0056] 2. TFD generation process

[0057] The trajectory flow graph is a two-dimensional grid matrix with CAV trajectories and SVs positions, such as Figure 1 As shown in the figure, the grids occupied by obstacles are uniformly set to 1. For the grids occupied by vehicles, their values ​​are set to 1; otherwise, their values ​​are 0. At the same time, the driving intention of the vehicle is characterized by the predicted trajectory. There are two main reasons: first, any driving intention will eventually be reflected in the driving trajectory, that is, the predicted trajectory has the general characteristics of all types of driving intentions (such as acceleration, deceleration, lane change, etc.); second, the change behavior essentially leads to right-of-way conflicts between vehicles, and the predicted trajectory directly reflects the next right-of-way demand of the vehicle. The use of trajectory flow graphs can promote the ability of neural networks to learn to negotiate right of way similar to drivers.

[0058] Specifically, in order to encode the spatiotemporal information of the predicted trajectory into a two-dimensional grid matrix, the value of the vehicle's position at time t is encoded as 1 and decreases linearly to 0 over time, that is, at time t+T D The value of the position is 0, where T D is the time length of the trajectory. This encoding allows the vehicle to make more fine-grained reasoning about the driving behaviors of surrounding vehicles and make more accurate decisions.

[0059] 3. Transformer network based on ResNet

[0060] Compared with convolutional neural network (CNN), residual network (ResNet) has a deeper model. It solves the gradient vanishing problem in deep network training through residual connection, allowing very deep networks to be trained, thereby capturing more layers of features. Transformer has a powerful global feature modeling capability and can capture more complex dependencies and long-distance information. When dealing with complex tasks such as lane change decisions, the Transformer network combined with ResNet can better capture the complex relationship between historical data such as vehicle speed, trajectory, steering angle, etc., thereby improving the decision-making performance of the model. Figure 3 As shown, the main steps of the algorithm are as follows:

[0061] (1) DFP feature extraction: After the RSU receives the DFP of CAV and SVs, the image is resized by bilinear interpolation to fit the input of ResNet-50. The resized image first passes through the initial convolution layer of ResNet-50 for feature extraction, then is normalized by the batch normalization layer, and the ReLU activation function is applied to introduce nonlinear characteristics to enhance the feature representation capability. Next, the image enters the residual block stacking part. Each residual block contains multiple convolution layers, batch normalization layers, and ReLU activation functions. The input features are retained through jump connections and training is accelerated. High-level features are gradually extracted through multiple residual blocks. After passing through the global average pooling layer, the feature map is flattened into a one-dimensional vector.

[0062] In the Transformer network, the feature vector extracted by ResNet-50 is first mapped to the feature dimension required by the Transformer through a linear transformation, and position encoding is added to incorporate the position information of the sequence. Specifically, the position is encoded using the following formula:

[0063]

[0064]

[0065] Where pos represents the position index of each element after DFP is flattened, i represents the dimension index of each element, and d mod Represents the dimension of the model. Subsequently, these feature vectors enter the multi-head self-attention mechanism to generate a query vector Q, a key vector K, and a value vector V with dimensions, and calculate the similarity between them to capture the relationship between features:

[0066]

[0067] By weighted summing these value vectors, the output of each attention head is obtained, and then the output of all heads is linearly transformed to generate the final multi-head self-attention output. Next, it is input into the Feedforward Neural Network (FFN), which consists of two linear transformation layers and one nonlinear activation function, which can be expressed as:

[0068] FFN(X)=W2σ(W1X) (4)

[0069] Among them, W1 and W2 are two parameter matrices of two linear transformation layers, and σ represents a nonlinear activation function. Then the training is stabilized by layer normalization, and the final output is:

[0070] LayerNorm(X+Attention(X))(5)

[0071] Among them, X is the input of the self-attention layer, and the query, key and value matrices Q, K and V are all derived from the same input matrix X. Finally, the output feature vector of the Transformer network is concatenated with the feature vector extracted by ResNet-50 and then sent to the fully connected layer for fusion with other features.

[0072] (2) TFD feature extraction: When the RSU receives the TFD of CAV and SVs, it inputs it into the CNN. First, it passes through a convolution layer containing 16 3×3 convolution kernels to perform a convolution operation on the input image to extract low-dimensional features. Then it passes through a 2×2 maximum pooling layer to reduce the size of the feature map while retaining the most important information. Next, it passes through a convolution layer containing 32 3×3 convolution kernels to further convolve the pooled feature map to extract high-dimensional features. It passes through a 2×2 maximum pooling layer again to further reduce the size of the feature map. Subsequently, the output of the pooling layer is flattened into a one-dimensional vector and input into a fully connected layer containing 64 neurons. It is processed through the ReLU activation function and finally outputs a feature vector containing 32 neurons. This feature vector is input into the fully connected layer and connected with the vector processed by the DFP through the neural network, and finally acts on the lane change decision.

[0073] (3) Traffic factor setting: The motivation of CAV to change lanes is to obtain faster speed and better driving conditions. At the same time, safety is the premise of lane change decision. These factors are jointly determined by adjacent vehicles. Therefore, the traffic factor is set as the weighted sum of incentive factors, safety factors and tolerance factors. j ,v k Expressed as the speed of CAV and SVs, d jk It is represented as the distance from vehicle j to the surrounding vehicle k. If there is no such vehicle around vehicle j, that is, Then v k and d jk will be set to 0. Where k∈{B,P,PL,FL,FR,ASL,ASR}, d max is the maximum safe distance on the current road. B, P, PL, FL, FR, ASL and ASR represent the following vehicle in the current lane, the leading vehicle in the current lane, the leading vehicle in the left adjacent lane, the leading vehicle in the right adjacent lane, the following vehicle in the left adjacent lane, the following vehicle in the right adjacent lane, the parallel vehicle in the left adjacent lane and the parallel vehicle in the right adjacent lane, respectively.

[0074] Motivational factors: Motivational factors include speed gain factors and headway gain factors. The speed gain obtained by the vehicle can be represented by the difference between the maximum achievable speed of the current lane and the maximum achievable speed of the target lane relative to the current speed; similarly, the headway gain can be represented by the difference between the current headway and the headway in the target lane. Since the target lane can be the left adjacent lane, the right adjacent lane or the current lane, the motivational factor can be represented as:

[0075] F i =f I ((v j -v P ),(v PL -v P ),(v PR -v P ),(d jPL -d jP ),(d jPR -d jP )) (6)

[0076] Safety factor: To ensure the safety of lane change, the risk of collision with the following vehicle in the target lane must be evaluated, which depends on the distance and speed difference between the current vehicle and the following vehicle. When the distance to the following vehicle in the target lane is large enough for the current vehicle to complete the lane change action, and the speed of vehicle j is faster than that of the following vehicle, the collision risk will be low. Therefore, the safety factor can be expressed as:

[0077] F s =f S ((d jFL ,d jFR ,(v j -v jFL ),(v j -v jFR )) (7)

[0078] Tolerance factor: The tolerance factor is a criterion used to determine whether the current lane is suitable for continued driving. It takes into account the distance between vehicle j and the vehicle in front and the safety time interval t h . For different roads, due to different speed limits, the safe distance between vehicles is also different. Although there is a minimum safe distance requirement, many drivers still choose to maintain a longer distance to ensure that they have more reaction time and drive more comfortably and safely. The tolerance factor helps the vehicle decide whether it should continue to drive in the current lane or change lanes to a more suitable lane, which can be expressed as:

[0079] F r =f R (d jP -v j ·t h) (8)

[0080] (4) Decision model: In order to formulate the lane change decision model, the incentive factors, safety factors, tolerance factors, historical driving features extracted from DFP and trajectory prediction information extracted from TFD of lane change are considered as follows:

[0081]

[0082] where y is a lane change decision vector containing three elements, corresponding to the probabilities of lane keeping, left lane change, and right lane change, and f DS (D) is the driving style extracted from DFP, f DT (T) is the trajectory information extracted from TFD, f θ The role of is to find the solution of the decision model. Since lane change decision is a multi-parameter, nonlinear problem, it is difficult to solve it using traditional mathematical methods. The present invention uses a deep learning model to solve this problem. Through the trajectory data set, the vehicle can extract input variables frame by frame to make decisions, and then model it as a supervised learning problem. The learning goal is to find f θ (·) Minimize the long-term average loss, defined as follows:

[0083]

[0084] Where N is the number of input-output pairs used for training, is the true decision vector of vehicle j obtained from the dataset, and the cross entropy function is used as the loss function L. If the actual decision is c, then otherwise

[0085] 4. System Process

[0086] Figure 4 The figure shows the execution flow chart of multi-dimensional cooperative vehicle lane change decision based on user portrait. The specific steps are as follows:

[0087] Steps 401-404: initialization phase;

[0088] Step 401: Algorithm initialization;

[0089] Step 402: CAV downloads edge decision model from RSU;

[0090] Step 403: The CAV collects local sensor data and generates a DFP;

[0091] Step 404: The CAV receives the DFP and TFP of the SVs;

[0092] Steps 405-408: CAV local driving decision stage;

[0093] Step 405: The CAV uses its own DFP and SVs data to make a lane change decision based on the edge model;

[0094] Step 406: The CAV generates its own TFP according to the decision action made;

[0095] Step 407: CAV shares its own DFP and TFP with SVs;

[0096] Step 408: The CAV makes a lane change decision and uploads the DFP and TFP to the experience replay pool;

[0097] Steps 409-412: edge model update phase;

[0098] Step 409: RSU performs experience replay pool sampling;

[0099] Step 410: The RSU processes the DFP and TFP into vectors through different networks and sends them to the fully connected layer;

[0100] Step 411: updating the model by combining the 10-dimensional vector of traffic factors;

[0101] Step 412: The algorithm ends.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-dimensional cooperative vehicle lane change decision method based on user portrait, characterized by: The method comprises the following steps: S1: Establish a data sharing model based on CAV-SVs-RSU; S2: CAV DFP generation based on historical driving data; S3: Lane change decision based on DFP and TFD of CAVs and SVs; S4: TFD generation based on lane-changing maneuvers of CAV; S5: DFP and TFD sharing of CAV; S6: Model training based on ResNet and Transformer.

2. The multi-dimensional cooperative vehicle lane change decision method based on user portrait according to claim 1 is characterized by: In S1, a data sharing model based on CAV-SVs-RSU is adopted to improve the safety of lane change decision by sharing data among CAV, SVs and RSU; CAV and SVs share DFP and TFD to improve CAV's perception of the surrounding environment.

3. The multi-dimensional collaborative vehicle lane change decision method based on user portrait according to claim 2 is characterized by: In S2, a two-dimensional matrix is ​​generated by calculating M statistical functions of vehicle driving characteristics, and the data is converted into DFP using time window extraction; specific features are extracted from historical driving information to implicitly describe the user profile of driving characteristics, the purpose of which is to perform lane change decision tasks and include more specific driving style information; the specific process of CAV converting historical data into DFP is as follows: (1) Data collection: The current vehicle j collects the relative position y through sensors. j , relative position x j , lateral speed Longitudinal speed Lateral acceleration Longitudinal acceleration Headway d jk 、Headway h jk etc. L historical driving features; where k∈{B,P,PL,FL,FR,ASL,ASR}; B, P, PL, FL, FR, ASL and ASR represent the following vehicle in the current lane, the preceding vehicle in the current lane, the preceding vehicle in the left adjacent lane, the preceding vehicle in the right adjacent lane, the following vehicle in the left adjacent lane, the following vehicle in the right adjacent lane, the parallel vehicles in the left adjacent lane and the parallel vehicles in the right adjacent lane, respectively; (2) Statistical function calculation: For the L driving characteristics, M statistical functions such as the mean value, standard deviation, median, 25% percentile, 75% percentile, minimum value, and maximum value are calculated for each second, and a column of a two-dimensional matrix is ​​generated based on the obtained results, that is, a column vector of size L·M×1 is generated every second, (L driving characteristics×M statistical functions); (3) Time window setting: As time accumulates, the two-dimensional matrix gradually becomes larger, reflecting the long-term driving characteristics of the CAV. Since the lane change time is very short, we focus on short-term driving characteristics and set a time window of length S seconds to frame the two-dimensional matrix S seconds before the lane change. The size of each DFP is L·M×S. (4) Normalization: Normalize all data in the two-dimensional matrix to eliminate the impact of different feature dimensions on DFP; (5) DFP generation: All data in the two-dimensional matrix are represented as different colors. The depth of the color reflects the size of the data, and the type of color represents the category of the driving feature. A driving feature map is generated, and the driving features are implicitly represented as an image.

4. The multi-dimensional cooperative vehicle lane change decision method based on user portrait according to claim 3 is characterized by: In said S4, the trajectory flow graph is a two-dimensional grid matrix with CAV trajectories and SVs positions, and the values ​​for grids occupied by obstacles are uniformly set to 1; for grids occupied by vehicles, the values ​​are set to 1; Otherwise, its value is 0; at the same time, the driving intention of the vehicle is represented by the predicted trajectory; In order to encode the spatiotemporal information of the predicted trajectory into a two-dimensional grid matrix, the value of the vehicle's position at time t is encoded as 1 and decreases linearly to 0 over time, that is, at time t+T D The value of the position is 0, where T D is the time length of the trajectory.

5. The multi-dimensional cooperative vehicle lane change decision method based on user portrait according to claim 4 is characterized by: In S6, the steps of combining the Transformer network with ResNet are as follows: (1) DFP feature extraction: After the RSU receives the DFP of CAV and SVs, the image is resized by bilinear interpolation to fit the input of ResNet-50. The resized image is first passed through the initial convolutional layer of ResNet-50 for feature extraction, and then normalized by the batch normalization layer. The ReLU activation function is applied to introduce nonlinear characteristics and enhance the feature representation capability. Next, the image enters the residual block stacking part. Each residual block contains multiple convolutional layers, batch normalization layers, and ReLU activation functions. The input features are retained through jump connections and the training is accelerated. High-level features are gradually extracted through multiple residual blocks. After passing through the global average pooling layer, the feature map is flattened into a one-dimensional vector. In the Transformer network, the feature vector extracted by ResNet-50 is first mapped to the feature dimension required by the Transformer through a linear transformation, and position encoding is added to incorporate the position information of the sequence. The position is encoded using the following formula: Where pos represents the position index of each element after DFP is flattened, i represents the dimension index of each element, and d mod Represents the dimension of the model; the feature vector enters the multi-head self-attention mechanism to generate a query vector Q, a key vector K, and a value vector V with dimensions, and calculates the similarity between them to capture the relationship between features: By weighted summing these value vectors, the output of each attention head is obtained, and then the output of all heads is linearly transformed to generate the final multi-head self-attention output; then, it is input into the Feedforward Neural Network (FFN), which consists of two linear transformation layers and one nonlinear activation function, expressed as: FFN(X)=W2σ(W1X) (4) Among them, W1 and W2 are two parameter matrices of two linear transformation layers, and σ represents a nonlinear activation function. Then the training is stabilized by layer normalization, and the final output is: LayerNorm(X+Attention(X))(5) Among them, X is the input of the self-attention layer, and the query, key and value matrices Q, K and V are all derived from the same input matrix X; Finally, the output feature vector of the Transformer network is concatenated with the feature vector extracted by ResNet-50, and then sent to the fully connected layer for fusion with other features; (2) TFD feature extraction: When the RSU receives the TFD of CAV and SVs, it inputs it into the CNN. First, it passes through a convolution layer containing 16 3×3 convolution kernels to perform a convolution operation on the input image to extract low-dimensional features. Then it passes through a 2×2 maximum pooling layer to reduce the size of the feature map while retaining the most important information. Next, it passes through a convolution layer containing 32 3×3 convolution kernels to further convolve the pooled feature map to extract high-dimensional features. It passes through a 2×2 maximum pooling layer again to further reduce the size of the feature map. Subsequently, the output of the pooling layer is flattened into a one-dimensional vector and input to a fully connected layer containing 64 neurons. It is processed through the ReLU activation function and finally outputs a feature vector containing 32 neurons. This feature vector is input into the fully connected layer and connected with the vector processed by the DFP through the neural network. Finally, it acts on the lane change decision. (3) Traffic factor setting: The motivation for CAV to change lanes is to obtain a faster speed and better driving conditions. At the same time, safety is the premise of lane change decision. These factors are jointly determined by adjacent vehicles. The traffic factor is set as the weighted sum of incentive factors, safety factors, and tolerance factors. j ,v k Expressed as the speed of CAV and SVs, d jk It is represented as the distance from vehicle j to the surrounding vehicle k. If there is no such vehicle around vehicle j, that is, Then v k and d jk will be set to 0; where k∈{B,P,PL,FL,FR,ASL,ASR}, d max is the maximum safe distance on the current road; B, P, PL, FL, FR, ASL and ASR represent the following vehicle in the current lane, the leading vehicle in the current lane, the leading vehicle in the left adjacent lane, the leading vehicle in the right adjacent lane, the following vehicle in the left adjacent lane, the following vehicle in the right adjacent lane, the parallel vehicles in the left adjacent lane and the parallel vehicles in the right adjacent lane respectively; Motivational factors: Motivational factors include speed gain factors and headway gain factors. The speed gain obtained by the vehicle is represented by the difference between the maximum achievable speed of the current lane and the maximum achievable speed of the target lane relative to the current speed. The headway gain is represented by the difference between the current headway and the headway in the target lane. The target lane is the left adjacent lane, the right adjacent lane or the current lane. The motivational factors are represented as follows: F i =f I ((v j -v P ),(v PL -v P ),(v PR -v P ),(d jPL -d jP ),(d jPR -d jP )) (6) Safety factor: evaluates the risk of collision with the following vehicle in the target lane, which depends on the distance and speed difference between the current vehicle and the following vehicle; when the distance to the following vehicle in the target lane is large enough for the vehicle to complete the lane change action, and the speed of vehicle j is faster than that of the following vehicle, the collision risk will be low; the safety factor is expressed as: F s =f S ((d jFL ,d jFR ,(v j -v jFL ),(v j -v jFR )) (7) Tolerance factor: The tolerance factor is a criterion used to determine whether the current lane is suitable for continued driving; it takes into account the distance between vehicle j and the vehicle in front and the safety time interval t h ; The tolerance factor helps the vehicle decide whether it should continue driving in the current lane or change to a more suitable lane, expressed as: F r =f R (d jP -v j ·t h ) (8) (4) Decision model: To develop a lane change decision model, the incentive factors, safety factors, tolerance factors, historical driving features extracted from DFP, and trajectory prediction information extracted from TFD are considered as follows: where y is a lane change decision vector containing three elements, corresponding to the probabilities of lane keeping, left lane change, and right lane change, and f DS (D) is the driving style extracted from DFP, f DT (T) is the trajectory information extracted from TFD, f θ The role of is to find the solution of the decision model; through the trajectory data set, the vehicle extracts input variables frame by frame to make decisions, and then models it as a supervised learning problem; the learning goal is to find f θ (·) Minimize the long-term average loss, defined as follows: Where N is the number of input-output pairs used for training, is the true decision vector of vehicle j obtained from the dataset, where the cross entropy function is used as the loss function L; if the actual decision is c, then otherwise

Images