An intelligent generation method for dynamic traffic scenes at road intersections

Abstract

The present invention discloses a method for intelligently generating dynamic traffic scenes at road intersections, comprising the following steps: constructing an initial lane topology map containing only lanes for an initial static map; utilizing the inverse diffusion process of GDSS to generate an initial heterogeneous map containing lanes and traffic participants based on the initial lane topology map, and constructing initial historical trajectory data based on the initial heterogeneous map; in each iteration, performing rasterization modeling on the current static map with the target traffic participant as the center to obtain the current rasterized map, utilizing a trajectory generation model constructed based on LSTM to generate future trajectory data corresponding to the current iteration based on the current historical trajectory data and the current rasterized map. Given a static map of the intersection, the method utilizes a trajectory generation model based on a neural network to rapidly, efficiently, and cost-effectively construct dynamic traffic scenes at the intersection, while reducing reliance on user control.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation technology, and in particular relates to a method for intelligently generating dynamic traffic scenes at road intersections. Background Art

[0002] Methods for automatically generating dynamic traffic scenarios primarily include procedural modeling, cellular automation, and data-driven generation. Procedural modeling, which uses mathematical models and basic rules to control the movement patterns of vehicles, offers low cost and high efficiency, but lacks diversity. Taking VISSIM as an example, the software pre-sets rules for various traffic scenarios, such as following a vehicle, changing lanes, and conflict detection. This allows users to easily generate relatively reasonable dynamic traffic scenarios, but diversity is difficult to ensure, making it difficult to meet the scenario coverage requirements required for autonomous driving evaluation.

[0003] A cellular automaton is a finite automaton system that provides a discrete computational model for the behavior of complex behavioral systems. This approach simulates dynamic traffic scenarios using cellular automata. This approach discretizes a static map into cells, abstracts traffic participants into integer cell sizes, defines a time step, and defines update rules to achieve the simulation of dynamic traffic scenarios. The cellular automaton model can represent the interactions between vehicles, allowing different vehicles to have different driving rules (acceleration / deceleration, lane change rules, reaction time), and define different rules for each vehicle. While this approach enhances the diversity of generated scenarios to a certain extent, it is still essentially a rule-based procedural modeling approach that cannot meet the needs.

[0004] The most crucial component of dynamic traffic scenarios is the motion patterns of traffic participants, such as vehicles and pedestrians. In current mainstream dynamic scenario datasets, such as NGSIM and INTERACTION, the motion of these participants is represented in trajectory format, recording the position, velocity, and other information of these participants at each point in time. Therefore, dynamic traffic scenario data can be viewed as a composition of the trajectories of different traffic participants. The simulation process of dynamic traffic scenarios can be considered the generation of trajectories for these participants. Therefore, data-driven dynamic traffic scenario simulation techniques are categorized into two types: direct vehicle trajectory generation and iterative vehicle trajectory generation. Direct vehicle trajectory generation techniques primarily utilize generative networks to directly generate different types of trajectories. In iterative trajectory generation techniques, the model specifies a time step during the iteration process, generating fixed (or variable) length vehicle trajectories each time, thus achieving iterative generation. However, the environmental factors involved in vehicle driving are complex and variable, making it impossible to comprehensively define them using simple rules. Therefore, deep learning and extensive datasets are required for training and learning to achieve motion modeling of each traffic participant in dynamic traffic scenario simulation. For example, J. Whale et al. proposed a data-driven generation method based on generative adversarial networks that focuses on specific vehicle trajectories, such as lane changes and overtaking. However, this method struggles to support large-scale scene generation and is difficult to adapt to complex scene generation tasks.

[0005] In summary, existing technologies mainly have problems such as insufficient diversity, difficulty in meeting the coverage requirements of autonomous driving evaluation, low efficiency or over-reliance on user control. Summary of the Invention

[0006] In view of the above, the object of the present invention is to provide a method for intelligently generating dynamic traffic scenes at road intersections to achieve efficient traffic trajectory prediction.

[0007] To achieve the above-mentioned purpose of the invention, an embodiment provides a method for intelligently generating a dynamic traffic scene at a road intersection, comprising the following steps:

[0008] Constructing an initial lane topology map containing only lanes for the initial static map;

[0009] The inverse diffusion process of the GDSS model is used to generate an initial heterogeneous graph containing lanes and traffic participants based on the initial lane topology graph, and the initial historical trajectory data is constructed based on the initial heterogeneous graph.

[0010] In each iteration, a rasterized model is constructed for the current static map centered on the target traffic participant to obtain the current rasterized map. The trajectory generation model constructed using a neural network generates future trajectory data corresponding to the current iteration based on the current historical trajectory data and the current rasterized map.

[0011] The rasterized map is updated according to the future trajectory data, and the future trajectory data is used as the historical trajectory data. The updated rasterized map is then combined with the updated rasterized map to participate in the prediction of the next iteration of future trajectory data.

[0012] Preferably, in the initial lane topology graph, lanes are abstracted as a type of lane node, and edges between lane nodes are constructed based on the connection relationship between lanes;

[0013] In the initial heterogeneous graph, lanes are abstracted as a type of lane node, and traffic participants are abstracted as participant nodes. Edges between lane nodes are constructed based on the connection relationship between lanes, and edges between participant nodes and lane nodes are constructed based on the relationship between traffic participants and lanes.

[0014] The attribute information of the lane node includes lane width, lane length, and lane type, wherein the lane type includes at least one of straight ahead, left turn, right turn, and U-turn;

[0015] The attribute information of the participant node includes the location, speed, and direction angle of the traffic participant.

[0016] Preferably, generating an initial heterogeneous graph based on the initial lane topology graph using the inverse diffusion process of the GDSS model includes:

[0017] Initialize the initial matrix G with random noise T (X T ,A T ), where X T Represents the node attribute matrix, A T Represents the adjacency matrix. The lane node attribute matrix and adjacency matrix of the initial lane topology graph are used as guidance information to replace part of the noise in the initial matrix to obtain the input data of the GDSS model. The GDSS model performs inverse diffusion based on the input data to obtain the initial heterogeneous graph.

[0018] Preferably, constructing initial historical trajectory data based on the initial heterogeneous graph includes:

[0019] According to the position, speed and orientation angle of the target traffic participant in the initial heterogeneous graph, the initial historical trajectory data of the target participant is obtained according to uniform linear motion.

[0020] Preferably, raster modeling is performed on the current static map with the target traffic participant as the center to obtain the current rasterized map, including:

[0021] A current rasterized map of M*N grids is established for the current static map with the target traffic participant as the center. The positions of other traffic participants in the current static map are adjusted in the rasterized map with the target traffic participant as the center point according to the adjustment rules. The adjustment rule is that each traffic participant can only occupy one grid, and there is only one traffic participant in a grid. Each grid is divided into two types of attributes: drivable and non-drivable according to the position of its center point in the current static map. The attribute of the drivable grid is set to 0, and the attribute of the non-drivable grid is set to -255.

[0022] Preferably, the trajectory generation model includes an LSTM encoder, a spatial feature extraction module and an LSTM decoder, wherein the LSTM encoder is used to encode according to historical trajectory data to extract trajectory temporal features, the spatial feature extraction module is used to place the trajectory temporal features into the positions of the corresponding traffic participants in the rasterized map, and after obtaining a feature map containing temporal features and position information, convolution and pooling are used to extract features from the feature map to obtain spatial features; the LSTM decoder is used to decode and linearly map the spatial features after fusing the temporal features to predict future trajectory data.

[0023] Preferably, the temporal features need to be linearly mapped before being fused with the spatial features, and the linearly mapped temporal features and spatial features are spliced ​​and fused.

[0024] Preferably, a trajectory generation model based on a neural network is used to generate future trajectory data corresponding to the current iteration based on the current historical trajectory data and the current rasterized map, including:

[0025] Use LSTM encoder to encode the current historical trajectory data to extract trajectory time series features;

[0026] The spatial feature extraction module is used to place the trajectory temporal features into the corresponding positions of traffic participants in the current rasterized map. After obtaining a feature map containing temporal features and position information, convolution and pooling are used to extract features from the feature map to obtain spatial features.

[0027] LSTM decoding is used to fuse spatial features with temporal features, and then decoded and linearly mapped to predict the future trajectory data corresponding to the current iteration of all traffic participants.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] Given static intersection map data, the system automatically generates the initial positions and motion patterns of traffic participants through inverse diffusion, reducing reliance on user control and ensuring the rapid, efficient, and low-cost construction of dynamic intersection traffic scenarios. Secondly, the dynamic intersection scene initialization process is transformed into a heterogeneous graph generation problem, allowing the model to focus on learning characteristics such as the initial speed and position of traffic participants. Finally, by iteratively generating the vehicle's future motion trajectory based on the rasterized map and the historical trajectory data of traffic participants, the system automatically generates dynamic traffic scenarios of indefinite length given a given initial state, ensuring the diversity of dynamic traffic scenarios to meet the coverage requirements of autonomous driving evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 This is a flow chart of a method for intelligently generating dynamic traffic scenes at road intersections provided by an embodiment;

[0032] Figure 2 The embodiment provides an initial static map (left) and a corresponding initial lane topology map (right);

[0033] Figure 3 The initial heterogeneous graph provided by the embodiment (left) and the corresponding visualization result (right);

[0034] Figure 4 is a schematic diagram of generating an initial heterogeneous graph based on a GDSS model provided in an embodiment;

[0035] Figure 5 Schematic diagram of GDSS model input data provided by the embodiment;

[0036] Figure 6 Schematic diagram of GDSS model output data provided by the embodiment;

[0037] Figure 7 This is a schematic diagram of drivable areas and non-drivable areas in a rasterized map provided by an embodiment;

[0038] Figure 8 Schematic diagram of the structure and reasoning of the trajectory generation model provided in the embodiment;

[0039] Figure 9 This is a schematic diagram of iteratively generating future trajectory data provided by an embodiment. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0041] like Figure 1 As shown, the embodiment provides a method for intelligently generating a dynamic traffic scene at a road intersection, comprising the following steps:

[0042] Step 1: Build an initial lane topology graph containing only lanes for the initial static map.

[0043] In the embodiment, the initial static map is as follows Figure 2 As shown in the middle left figure, it only contains lane information. Based on this initial static map, Figure 2 The initial lane topology shown in the middle right figure shows lanes abstracted as lane nodes. Edges between lane nodes are constructed based on the connection relationships between lanes. Lane node attribute information includes lane width, lane length, and lane type. Lane type can include at least one of straight ahead, left turn, right turn, and U-turn. Table 1 shows an example of the attribute information for a lane node:

[0044] Table 1

[0045] Lane width Lane length straight Turn left Turn right U-turn 1.0 1.2 1 0 1 0

[0046] Step 2: Generate an initial heterogeneous graph based on the initial lane topology graph using the inverse diffusion process of the GDSS model.

[0047] In an embodiment, the GDSS model adopts the idea of ​​a diffusion generative model. By training and learning the GDSS model, a mapping from the noise space to the real sample data space is obtained, so that the noise data can be generated into a real sample through the mapping. The GDSS model includes a forward diffusion process and a reverse diffusion process. First, through the forward diffusion process, noise data is continuously and smoothly added to the real sample data set to obtain a continuous and smooth transition data set from the real sample data to the Gaussian noise data; then, in the reverse diffusion process, the training goal is defined as the Euclidean distance between the generated data and the real sample data is as small as possible, that is, the generated data is as close to the real sample data as possible; finally, the GDSS model finds the mapping from the Gaussian noise data space to the real data space through training, and can generate data in the real sample data space through the input Gaussian noise data, that is, generate new graph data.

[0048] The diffusion process of the GDSS model is expressed as the diffusion process can be expressed as *G t =(X t,A t )+ t ∈,0,T-, that is, in the forward diffusion process, noise G is gradually generated by adding noise data based on the input real data G0 T , in the reverse diffusion process based on the input G T The real data G0 is gradually restored by removing noise.

[0049] Since the goal of the dynamic traffic scene initialization generation problem is to generate the initial position, orientation and other information of traffic participants that conform to the static map, the attributes and topological relationships of lane nodes are not generated during the heterogeneous graph generation and reasoning process. Instead, they are used as guidance information. The model only generates the nodes of traffic participants, including their attribute information and topological relationships. Therefore, the GDSS model is improved, and the initial noise matrix G of the inverse diffusion is T The original lane information is added as guidance information, and then the trained GDSS can better generate traffic participant information based on lane information guidance.

[0050] Since the initial lane topology graph input into the GDSS model contains lane information, the generated initial heterogeneous graph contains both lane information and traffic participant information. Figure 3 As shown in the middle left figure, in this initial heterogeneous graph, lanes are abstracted as a type of lane node, and traffic participants are abstracted as participant nodes. Edges between lane nodes are constructed based on the connection relationship between lanes, and edges between participant nodes and lane nodes are constructed based on the relationship between traffic participants and lanes. The attribute information of participant nodes includes the location, speed, and heading angle of the traffic participant. Table 2 shows an example of the attribute information of a participant node:

[0051] Table 2

[0052] Traffic participant position (S coordinate) Traffic participant position (t coordinate) speed Facing angle 0.5 0.5 0.7 0.0

[0053] It should be noted that the initial heterogeneous graph includes the target traffic participant of interest, as well as other traffic participants around the target participant. Traffic participants refer to entities participating in traffic, including vehicles, pedestrians, etc.

[0054] Based on the modified GDSS model above, the inverse diffusion process of the GDSS model is used to generate the initial heterogeneous graph based on the initial lane topology graph, such as Figure 4 As shown, including: random noise initialization initial matrix G T =(X T ,A T ), where X T Represents the node attribute matrix, A TRepresents the adjacency matrix. The lane node attribute matrix and adjacency matrix of the initial lane topology map are used as guidance information to replace part of the noise in the initial matrix to obtain the input data of the GDSS model, such as Figure 5 As shown, the GDSS model performs reverse diffusion based on the input data to obtain the initial heterogeneous graph, as shown in Figure 6 shown.

[0055] Step 3: Construct the initial historical trajectory data based on the initial heterogeneous graph.

[0056] In the embodiment, after obtaining the initial heterogeneous graph, since the initial heterogeneous graph contains attribute information such as the position, speed, and heading angle of the target traffic participant, the initial historical trajectory data of the target participant is obtained according to the position, speed, and heading angle of the target traffic participant in accordance with uniform linear motion.

[0057] Step 4: In the current iteration, a rasterized model is performed on the current static map with the target traffic participant as the center to obtain the current rasterized map.

[0058] In dynamic scene datasets, roads are typically represented as centerline diagrams or closed graphs of drivable areas, making them difficult to directly use as model input and difficult to integrate with other trajectory and spatial information. Therefore, this paper proposes a static map rasterization modeling method to construct a rasterized map. This discretizes the static map while preserving road shape information, allowing the trajectory generation model to learn from the static map information.

[0059] In the embodiment, a current grid map of M*N grids is established for the current static map with the target traffic participant as the center. During the gridding process, the size of each grid needs to be defined. For example, the width W of each grid is g Equal to the standard width of the main road lane at the intersection, which is 3.5m; the length of each grid L g Equal to the standard length of a small car, which is 4m. During the rasterization process, the positions of other traffic participants in the current static map are adjusted in accordance with the adjustment rules with the target traffic participant as the center point. The adjustment rule is that each traffic participant can only occupy one grid, and there is only one traffic participant in one grid. According to the adjustment rule, when there are multiple traffic participants in a grid, the traffic participants with larger areas are retained in the original grid, and the traffic participants with smaller areas are adjusted to other adjacent grids. Each grid is divided into two types of attributes, drivable and non-drivable, according to the position of its center point in the current static map. The attributes of drivable grids are set to 0, and the attributes of non-drivable grids are set to -255. For example Figure 7 shown.

[0060] Step 5: Use the trajectory generation model built by the neural network to generate future trajectory data corresponding to the current iteration based on the current historical trajectory data and the current rasterized map.

[0061] In the embodiment, Figure 8 As shown, the trajectory generation model includes an LSTM (Long Short-Term Memory) encoder, a spatial feature extraction module, and an LSTM decoder. The LSTM encoder is used to encode historical trajectory data to extract temporal features of the trajectory. Since the LSTM encoder alone cannot obtain the positional relationship features of surrounding traffic participants other than the target traffic participant in the current frame, the vehicle relies not only on the historical trajectories of surrounding traffic participants for future movement planning, but also on the relative positions of surrounding traffic participants and the target traffic participant. Therefore, the extracted trajectory temporal features are placed in a gridded map according to the positions of the corresponding traffic participants in the grid. Based on this, the spatial feature extraction module is used to place the trajectory temporal features in the gridded map according to the positions of the corresponding traffic participants. After obtaining a feature map containing temporal features and position information, the feature map is extracted using convolution and pooling to obtain spatial features. The LSTM decoder is used to decode and linearly map the fusion of spatial and temporal features to predict future trajectory data. It should be noted that the temporal features need to undergo linear mapping before being fused with the spatial features, and the linearly mapped temporal features and spatial features are spliced ​​and fused.

[0062] The LSTM encoder uses a weight-sharing LSTM encoder, meaning each LSTM encoder encodes a traffic participant's historical trajectory data at a single moment to generate trajectory time series features. The LSTM decoder includes a weight-sharing LSTM and FC, meaning each LSTM and FC is used to predict the traffic participant's trajectory data at a future moment.

[0063] Based on the above trajectory generation model, the current historical trajectory data and the current rasterized map generate the future trajectory data corresponding to the current iteration, including the following steps:

[0064] Use the LSTM encoder to encode the current historical trajectory data (e.g., the trajectory data of the past 30 frames) to extract the trajectory time series features;

[0065] The spatial feature extraction module is used to place the trajectory temporal features into the corresponding positions of traffic participants in the current rasterized map. After obtaining a feature map containing temporal features and position information, convolution and pooling are used to extract spatial features from the feature map. The spatial features are linearly mapped through a fully connected layer and then concatenated with the spatial features to obtain fused features.

[0066] The fused features are decoded and linearly mapped using the LSTM decoder to predict the future trajectory data corresponding to the current iteration of all traffic participants, which are displayed by coordinates.

[0067] Step 6: Update the rasterized map based on the future trajectory data, use the future trajectory data as historical trajectory data, and combine the updated rasterized map to participate in the prediction of the next iteration of future trajectory data, that is, jump to step 4.

[0068] In this embodiment, generating future trajectory data is an iterative process, with steps 4 and 5 forming one iteration. After obtaining the future trajectory data corresponding to one iteration, the rasterized map is updated based on the future trajectory data. Specifically, the static map is first updated based on the future trajectory data, and then the rasterized map is updated by rasterizing the static map. It should be noted that the rasterized map of the first iteration is created based on the initial heterogeneous map. This embodiment also updates historical trajectory data based on the future trajectory data, for example, extracting 30 frames from the current moment from the future trajectory data as historical trajectory data for the next iteration.

[0069] like Figure 9 As shown, after updating the rasterized map and historical trajectory data, jump to step 4 and iterate to generate future trajectory data to construct a dynamic traffic scene of a road intersection of indefinite length.

[0070] The above-described embodiment provides a method for intelligently generating dynamic traffic scenes at road intersections. This method, through a data-driven approach, learns from existing dynamic traffic scene datasets, overcoming the lack of diversity and meeting the coverage requirements of autonomous driving evaluation. Furthermore, given a user-provided static map of the intersection, a neural network-based trajectory generation model is used to rapidly, efficiently, and cost-effectively construct dynamic traffic scenes at the intersection, reducing reliance on user control.

[0071] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for intelligently generating dynamic traffic scenes at road intersections, characterized in that: The following steps are involved: Construct an initial lane topology graph containing only lanes for the initial static map. In the initial lane topology graph, lanes are abstracted as a type of lane node. Edges between lane nodes are constructed based on the connection relationship between lanes. Lane node attribute information includes lane width, lane length, and lane type. Lane type can be at least one of straight, left turn, right turn, and U-turn. The inverse diffusion process based on the GDSS model and the initial lane topology map generates an initial heterogeneous graph containing lanes and traffic participants, including: random noise initialization of the initial matrix G T =(X T ,A T ), where X T Represents the node attribute matrix, A T Representing the adjacency matrix, the lane node attribute matrix and adjacency matrix of the initial lane topology graph are used as guidance information to replace some noise in the initial matrix to obtain the input data of the GDSS model. The GDSS model performs inverse diffusion based on the input data to obtain the initial heterogeneous graph; In the initial heterogeneous graph, lanes are abstracted as a type of lane node, and traffic participants are abstracted as participant nodes. Edges between lane nodes are constructed based on the connection relationship between lanes, and edges between participant nodes and lane nodes are constructed based on the relationship between traffic participants and lanes. The attribute information of participant nodes includes the location, speed, and heading angle of traffic participants. Construct initial historical trajectory data based on the initial heterogeneous graph; In each iteration, a rasterized model is constructed for the current static map centered on the target traffic participant to obtain the current rasterized map. The trajectory generation model constructed using a neural network generates future trajectory data corresponding to the current iteration based on the current historical trajectory data and the current rasterized map, including: The trajectory generation model includes an LSTM encoder, a spatial feature extraction module, and an LSTM decoder. The LSTM encoder is used to encode historical trajectory data to extract trajectory temporal features. The spatial feature extraction module is used to place trajectory temporal features into the positions of corresponding traffic participants in the rasterized map. After obtaining a feature map containing temporal features and position information, convolution and pooling are used to extract spatial features from the feature map. The LSTM decoder is used to decode and linearly map the spatial features after fusing them with temporal features to predict future trajectory data. Use LSTM encoder to encode the current historical trajectory data to extract trajectory time series features; The spatial feature extraction module is used to place the trajectory temporal features into the corresponding positions of traffic participants in the current rasterized map. After obtaining a feature map containing temporal features and position information, convolution and pooling are used to extract features from the feature map to obtain spatial features. LSTM decoding is used to fuse spatial features with temporal features, and then decode and linearly map them to predict the future trajectory data of all traffic participants corresponding to the current iteration; The rasterized map is updated according to the future trajectory data, and the future trajectory data is used as the historical trajectory data. The updated rasterized map is then combined with the updated rasterized map to participate in the prediction of the next iteration of future trajectory data.

2. The method for intelligently generating dynamic traffic scenes at road intersections according to claim 1, characterized in that: Construct the initial historical trajectory data based on the initial heterogeneous graph, including: According to the position, speed and orientation angle of the target traffic participant in the initial heterogeneous graph, the initial historical trajectory data of the target participant is obtained according to uniform linear motion.

3. The method for intelligently generating dynamic traffic scenes at road intersections according to claim 1, characterized in that: The current static map is rasterized and modeled with the target traffic participant as the center to obtain the current rasterized map, including: A current rasterized map of M*N grids is established for the current static map with the target traffic participant as the center. The positions of other traffic participants in the current static map are adjusted in the rasterized map with the target traffic participant as the center point according to the adjustment rules. The adjustment rule is that each traffic participant can only occupy one grid, and there is only one traffic participant in a grid. Each grid is divided into two types of attributes: drivable and non-drivable according to the position of its center point in the current static map. The attribute of the drivable grid is set to 0, and the attribute of the non-drivable grid is set to -255.

4. The method for intelligently generating dynamic traffic scenes at road intersections according to claim 1, characterized in that: The temporal features need to undergo linear mapping before being fused with the spatial features. The linearly mapped temporal features and spatial features are then spliced ​​and fused.

Images