Conflict domain detection method and system in multi-vehicle conflict situation at unsignalized intersection
By identifying multi-vehicle conflict domains at unsignalized intersections using a 3D spatiotemporal grid map and a T-GCN spatiotemporal neural network model, the accuracy and adaptability issues of multi-vehicle conflict detection in dynamic environments in existing technologies are resolved, enabling efficient conflict risk early warning and management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-05-14
- Publication Date
- 2026-06-19
AI Technical Summary
Existing collision detection methods at unsignalized intersections lack adaptability in dynamic environments, the ability to handle dense targets, and the ability to generalize across scenarios. They are unable to effectively identify the complexity of multi-vehicle collision situations, resulting in high false alarm and false negative rates.
Using a 3D spatiotemporal grid map and a T-GCN spatiotemporal graph neural network model, we predict future vehicle trajectory distribution by constructing an enhanced dataset, identify potential conflict events and aggregate them into conflict domains, identify conflict domains by combining temporal proximity and spatial overlap conditions, and analyze the characteristics of conflict domains using a depth-first search algorithm.
It improves the accuracy and foresight of collision detection at unsignalized intersections, dynamically identifies multi-vehicle collision domains, reduces false alarms and false negatives, and has good robustness and generalization ability, adapting to complex traffic flow dynamics and environmental changes.
Smart Images
Figure CN122245115A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of traffic conflict detection technology, and relates to a conflict domain detection method and system for multi-vehicle conflicts at unsignalized intersections. Background Technology
[0002] In urban road traffic systems, unsignaled intersections, lacking traffic light control, rely primarily on driver judgment and traffic rules for the passage of vehicles, pedestrians, and non-motorized vehicles. However, with increasing traffic flow and the development of autonomous driving technology, traditional manual driving practices are insufficient to guarantee the safety and efficiency of unsignaled intersections, often leading to conflicts and accidents. Therefore, it is necessary to detect and analyze potential conflicts at intersections, thereby guiding drivers to avoid conflicting traffic flows and reducing the accident rate.
[0003] To guide drivers to avoid conflicting traffic flows, accurate detection and analysis of potential conflicts at intersections are crucial. Especially when multi-directional traffic flows converge, multiple conflict points are concentrated in the same spatiotemporal area, forming a conflict domain. In this case, the conflict risk is far greater than in the case of a single conflict point. Therefore, it is necessary to determine the spatiotemporal impact range of each independent conflict point in multi-vehicle conflict scenarios, aggregating conflict points within a certain range into a conflict domain for analysis. Currently, methods for intersection conflict detection can be categorized into geometric model-based methods, sensor-based methods, vehicle-road cooperative methods, and data-driven methods.
[0004] Geometric model-based methods assess conflict risk by predefined conflict points or by calculating the spatial overlap of vehicle trajectory envelopes, such as Time-to-Clash (TTC) models or trajectory intersection detection. Their advantages lie in their clear physical meaning and high computational efficiency, but their core drawback is their reliance on static conflict point assumptions, making them unable to adapt to dynamic traffic flow changes. Furthermore, they fail to consider the impact of environmental factors, leading to a higher false alarm rate in complex intersection scenarios.
[0005] Sensor-based methods utilize vehicle-mounted or roadside sensors to capture target states and achieve conflict warnings through multi-target tracking and trajectory prediction. Their advantage lies in acquiring high-precision local data, but their core drawbacks include limited sensing range due to occlusion, sensor performance degradation in extreme weather, and susceptibility to trajectory association errors in dense target scenarios.
[0006] Vehicle-to-everything (V2X) methods share vehicle pose and intent information through V2V / V2I communication and perform global conflict detection on an edge computing platform. Their advantage lies in overcoming the limitations of single-vehicle perception, but their core drawbacks include communication latency leading to decreased decision-making timeliness, perception blind spots for non-connected vehicles, and the risk of command loss due to channel congestion in high-density scenarios.
[0007] Data-driven approaches use machine learning models to learn historical trajectory patterns and predict future conflict probabilities. Their advantage lies in modeling complex interactive behaviors, but their core drawbacks include reliance on massive amounts of labeled data, the black-box nature of model decision-making hindering safety verification, and insufficient generalization ability across city intersections.
[0008] Traditional conflict detection methods rely on a single approach, failing to achieve a balance between adaptability to dynamic environments, the ability to handle dense targets, the insight into risk coupling, and the ability to generalize across scenarios. Therefore, there is an urgent need for a safety decision-making paradigm that moves beyond the traditional single-conflict-point paradigm and considers the complexity of multi-vehicle conflict scenarios. Summary of the Invention
[0009] This invention addresses the shortcomings of existing technologies by providing a method and system for collision domain detection in multi-vehicle collision scenarios at unsignalized intersections. The technical solution adopted by this invention is as follows:
[0010] A collision domain detection method for multi-vehicle collisions at an unsignalized intersection includes the following steps:
[0011] c1. Acquire historical trajectory data, real-time trajectory data, road structure information, and environmental information within unsignalized intersections. Discretize the spatial region of the intersection into a regular grid and construct a three-dimensional spatiotemporal raster map by combining it with the time dimension. Each grid cell includes vehicle attributes at the corresponding spatial location within the corresponding time segment.
[0012] c2. Based on the spatiotemporal grid map, the grid cell information and environmental information are fused to construct an enhanced dataset. Based on the enhanced dataset, a pre-trained T-GCN spatiotemporal graph neural network model is used to predict the probability distribution of the driving trajectory of all vehicles in the future period. The predicted trajectory is mapped back to the spatiotemporal grid map, and the probabilistic occupancy area of each vehicle in the future spatiotemporal space is marked.
[0013] c3. Based on the predicted vehicle occupancy information in the spatiotemporal grid map, and based on the preset conflict determination rules, identify potential conflict events between vehicles and extract the key features of each conflict event;
[0014] c4. Based on the correlation conditions of conflict events, analyze the spatiotemporal relationship between conflict events, identify conflict events that are coupled in spatiotemporal space, and aggregate the interrelated conflict events to form a conflict domain;
[0015] c5. For each conflict domain, analyze the corresponding set of conflict events to determine the corresponding conflict characteristics.
[0016] Furthermore, the historical trajectory data and real-time trajectory data include timestamps, longitude, latitude, speed, lateral acceleration, longitudinal acceleration, and heading angle; the road structure information includes lane boundaries and stop lines; and the environmental information includes weather and road surface friction coefficient.
[0017] Furthermore, the steps for constructing the three-dimensional spatiotemporal raster map include:
[0018] A spatiotemporal cube is constructed using a multidimensional raster layer to initialize a three-dimensional raster data structure. The dimensions include the number of spatial columns (X-axis), the number of spatial rows (Y-axis), and the number of time segments (T-axis).
[0019] The spatial area of the intersection is divided into a uniform grid, and the observation period is divided into continuous time segments;
[0020] The timestamp and coordinate information of the vehicle trajectory are mapped to the corresponding time segments and spatial grids, and the vehicle attributes are updated to the corresponding spatiotemporal grid cells.
[0021] Furthermore, based on the augmented dataset, a pre-trained T-GCN spatiotemporal graph neural network model is used to predict the probability distribution of the driving trajectories of all vehicles over a future period. Specific steps include:
[0022] The spatiotemporal raster graph is transformed into a spatiotemporal graph structure, which includes a vertex set, an edge set, and an adjacency matrix. Each vertex corresponds one-to-one with a raster cell, and the vertex features are defined by the enhanced feature data of the corresponding raster cell. The adjacency matrix is constructed using Gaussian radial basis functions based on the distance between the raster cells corresponding to the vertices.
[0023] The spatiotemporal graph structure is input into a pre-trained T-GCN spatiotemporal graph neural network model. Spatial features are aggregated through graph convolutional layers, and temporal dependencies are learned through gated recurrent units to output hidden states that fuse spatiotemporal features. Based on the hidden states at the final moment, the probability distribution of the driving trajectories of all vehicles in the future is generated.
[0024] Furthermore, the specific steps for identifying potential conflict events between vehicles and extracting key features of each conflict event include:
[0025] Traverse all vehicle pairs, calculate the joint occupancy probability of each vehicle pair within a spatiotemporal grid cell, and identify spatiotemporal grid cells whose joint occupancy probability exceeds a preset threshold as a potential conflict point of the vehicle pair.
[0026] For each vehicle pair, a conflict point graph is constructed based on all its potential conflict points; spatiotemporal connectivity analysis is performed on the conflict point graph, and all potential conflict points belonging to the same connectivity are aggregated into an independent conflict event.
[0027] For each independent conflict event, key features are extracted; the key features include at least one or more of the following: the vehicle pair involved in the conflict, the time window of the conflict, the spatial location of the conflict, and the severity of the conflict.
[0028] Furthermore, the association conditions for conflict events include temporal proximity and spatial overlap. Two conflict events are determined to be associated if and only if both temporal proximity and spatial overlap conditions are met simultaneously.
[0029] Furthermore, the specific steps for identifying a set of conflicting events that are coupled in space and time include:
[0030] Construct an association graph with each conflict event as a node. If two conflict events are associated, add an undirected edge between the corresponding nodes.
[0031] A depth-first search algorithm is used to perform connectivity component analysis on the constructed association graph, and each maximal connected subgraph obtained from the analysis is defined as a conflict domain.
[0032] Furthermore, the conflict characteristics include at least one of conflict vehicle set, spatiotemporal coverage, spatiotemporal features, and conflict coupling strength; the spatiotemporal features include at least one of centroid location, spatial density, spatiotemporal density, and spatiotemporal heat map; the conflict coupling strength includes at least one of average conflict strength, maximum cascading conflict, and conflict strength index.
[0033] A collision domain detection system for multi-vehicle collisions at an unsignalized intersection, used to implement the above method, includes:
[0034] Data extraction module: used to acquire historical trajectory data, real-time trajectory data, road structure information and environmental information within unsignalized intersections, discretize the spatial area of the intersection into a regular grid, and construct a three-dimensional spatiotemporal raster map by combining the time dimension; where each raster cell includes the vehicle attributes of the corresponding spatial location within the corresponding time segment;
[0035] The trajectory prediction module is used to construct an enhanced dataset by fusing grid cell information and environmental information based on the spatiotemporal grid map. Based on the enhanced dataset, it uses a pre-trained T-GCN spatiotemporal graph neural network model to predict the probability distribution of the driving trajectory of all vehicles in the future period of time, maps the predicted trajectory back to the spatiotemporal grid map, and marks the probabilistic occupancy area of each vehicle in the future spatiotemporal space.
[0036] Conflict identification module: Based on the predicted vehicle occupancy information in the spatiotemporal grid map and a preset conflict determination rule, it identifies potential conflict events between vehicles.
[0037] Conflict Correlation Module: Based on the correlation conditions of conflict events, it analyzes the spatiotemporal relationships between conflict events, identifies conflict events that are coupled in spatiotemporal space, and aggregates the interrelated conflict events to form a conflict domain;
[0038] Feature Analysis Module: Used to analyze each conflict domain based on the corresponding set of conflict events to determine the corresponding conflict features.
[0039] A computer device, the computer device comprising:
[0040] One or more processors;
[0041] Memory, used to store one or more programs;
[0042] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described collision domain detection method for multi-vehicle collisions at unsignalized intersections.
[0043] Compared with the prior art, the beneficial effects of the present invention are:
[0044] First, this invention improves the accuracy and foresight of conflict detection at unsignalized intersections. By using dynamically updated spatiotemporal grid maps and T-GCN spatiotemporal neural networks for trajectory prediction, it can effectively perceive complex traffic flow dynamics and vehicle-to-vehicle interactions, accurately capturing potential future conflict risks. Compared to traditional methods relying on static models or single sensors, this solution is more adaptable to environmental changes and dense traffic flow, significantly reducing false alarms and false negatives.
[0045] Secondly, this invention achieves a cognitive upgrade from single-point conflict to regional risk situation. The core of this cognitive upgrade lies in the proposed dynamic identification mechanism of "conflict domain." This mechanism can intelligently analyze and aggregate multiple conflict events that are closely related in time and space to form a high-risk area concept with clear time and space boundaries, i.e., a conflict domain. This breaks through the limitation of traditional methods that only focus on isolated conflict points, providing a more effective analytical dimension for understanding and responding to multi-vehicle conflict scenarios that influence each other within intersections, and greatly enhancing the comprehensiveness of risk warning.
[0046] Third, this invention possesses excellent robustness, generalization ability, and engineering implementation potential. The method employs a purely data-driven architecture, independent of specific hardware, effectively avoiding issues such as sensor obstruction, communication delays, or weather effects. Its design offers good interpretability, facilitates safety verification, and demonstrates adaptability to various intersection structures. This scheme boasts high computational efficiency and a relatively low deployment threshold, providing a practical and reliable technical foundation for the safe and efficient management of unsignalized intersections. Attached Figure Description
[0047] Figure 1 This is a schematic diagram of the overall method in an embodiment of the present invention.
[0048] Figure 2 This is a schematic diagram of trajectory spatiotemporal data prediction via G-TCN network in an embodiment of the present invention.
[0049] Figure 3 This is an example diagram of the spatiotemporal heat map of the conflict domain in an embodiment of the present invention. Detailed Implementation
[0050] The technical solution of the present invention will be further explained clearly and in detail below with reference to the accompanying drawings and specific examples.
[0051] Spatiotemporal raster maps, as a tool that integrates space, time, and specific attribute values into a unified data model for visualization and analysis, extend traditional two-dimensional spatial raster data to the temporal dimension. This invention, based on spatiotemporal raster maps, uses a T-GCN network to predict the sequence of spatiotemporal trajectory data, mapping the probability distribution of the predicted trajectory onto the spatiotemporal raster map for processing. Based on the spatiotemporal raster data, conflict events are detected and their correlation analyzed, thereby dynamically aggregating related conflict events into conflict domains and determining the boundaries of these conflict domains and their associated conflict event sets.
[0052] A collision domain detection method for multi-vehicle collisions at an unsignalized intersection includes the following steps:
[0053] c1. Acquire historical trajectory data, real-time trajectory data, road structure information, and environmental information within unsignalized intersections. Discretize the spatial region of the intersection into a regular grid and construct a three-dimensional spatiotemporal raster map by combining it with the time dimension; where each grid cell includes the vehicle attributes of the corresponding spatial location within the corresponding time segment.
[0054] The historical and real-time trajectory data include timestamps, longitude, latitude, speed, lateral acceleration, longitudinal acceleration, heading angle, etc.; road structure information includes lane boundaries, stop lines, etc.; and environmental information includes weather, road surface friction coefficient, etc.
[0055] The steps for constructing a 3D spatiotemporal raster map include:
[0056] A spatiotemporal cube is constructed using a multidimensional raster layer to initialize a three-dimensional raster data structure. The dimensions include the number of spatial columns (X-axis), the number of spatial rows (Y-axis), and the number of time segments (T-axis).
[0057] The spatial area of the intersection is divided into a uniform grid, and the observation period is divided into continuous time segments;
[0058] The timestamp and coordinate information of the vehicle trajectory are mapped to the corresponding time segments and spatial grids, and the vehicle attributes are updated to the corresponding spatiotemporal grid cells.
[0059] c2. Based on the spatiotemporal raster map, fused raster cell information with environmental information to construct an enhanced dataset. Using the enhanced dataset, a pre-trained T-GCN spatiotemporal graph neural network model was used to predict the probability distribution of the driving trajectories of all vehicles over a future period. The predicted trajectories were then mapped back to the spatiotemporal raster map, and the probabilistically occupied areas of each vehicle in the future spatiotemporal space were labeled. Specific steps include:
[0060] The spatiotemporal raster graph is transformed into a spatiotemporal graph structure, which includes a vertex set, an edge set, and an adjacency matrix. Each vertex corresponds one-to-one with a raster cell, and the vertex features are defined by the enhanced feature data of the corresponding raster cell. The adjacency matrix is constructed using Gaussian radial basis functions based on the distance between the raster cells corresponding to the vertices.
[0061] The spatiotemporal graph structure is input into a pre-trained T-GCN spatiotemporal graph neural network model. Spatial features are aggregated through graph convolutional layers, and temporal dependencies are learned through gated recurrent units to output hidden states that fuse spatiotemporal features. Based on the hidden states at the final moment, the probability distribution of the driving trajectories of all vehicles in the future is generated.
[0062] c3. Based on the predicted vehicle occupancy information in the spatiotemporal grid map, and according to preset conflict determination rules, identify potential conflict events between vehicles and extract key features for each conflict event. Specific steps include:
[0063] Traverse all vehicle pairs, calculate the joint occupancy probability of each vehicle pair within a spatiotemporal grid cell, and identify spatiotemporal grid cells whose joint occupancy probability exceeds a preset threshold as a potential conflict point of the vehicle pair.
[0064] For each vehicle pair, a conflict point graph is constructed based on all its potential conflict points; spatiotemporal connectivity analysis is performed on the conflict point graph, and all potential conflict points belonging to the same connectivity are aggregated into an independent conflict event.
[0065] For each independent conflict event, key features are extracted; the key features include at least one or more of the following: the vehicle pair involved in the conflict, the time window of the conflict, the spatial location of the conflict, and the severity of the conflict.
[0066] c4. Based on the correlation conditions of conflict events, analyze the spatiotemporal relationships between conflict events, identify conflict events that are coupled in spatiotemporal space, and aggregate the interrelated conflict events to form a conflict domain.
[0067] The association conditions for conflict events include temporal proximity and spatial overlap. Two conflict events are considered to be associated if and only if both temporal proximity and spatial overlap conditions are met simultaneously.
[0068] The specific steps for identifying a set of conflicting events that are coupled in space and time include:
[0069] Construct an association graph with each conflict event as a node. If two conflict events are associated, add an undirected edge between the corresponding nodes.
[0070] A depth-first search algorithm is used to perform connectivity component analysis on the constructed association graph, and each maximal connected subgraph obtained from the analysis is defined as a conflict domain.
[0071] c5. For each conflict domain, analyze the corresponding set of conflict events to determine the corresponding conflict characteristics. The conflict characteristics include at least one of the following: set of conflicting vehicles, spatiotemporal coverage, spatiotemporal features, and conflict coupling strength; the spatiotemporal features include at least one of the following: centroid location, spatial density, spatiotemporal density, and spatiotemporal heatmap; the conflict coupling strength includes at least one of the following: average conflict strength, maximum cascading conflict, and conflict intensity index. Figure 1 As shown, in one specific embodiment of the present invention, a collision domain detection method is provided for multi-vehicle collisions at an unsignalized intersection, specifically:
[0072] c1. Acquire historical trajectory data, real-time trajectory data, road structure information, and environmental information within unsignalized intersections. Discretize the spatial region of the intersection into a regular grid and construct a three-dimensional spatiotemporal raster map by combining it with the time dimension. Each grid cell records the vehicle attributes of the corresponding spatial location within the corresponding time segment, including vehicle occupancy status and motion characteristics.
[0073] c2. Based on the spatiotemporal grid map, the grid cell information and environmental information are fused to construct an enhanced dataset. Based on the enhanced dataset, a pre-trained T-GCN spatiotemporal graph neural network model is used to predict the probability distribution of the driving trajectory of all vehicles in the future period. The predicted trajectory is mapped back to the spatiotemporal grid map, and the probabilistic occupancy area of each vehicle in the future spatiotemporal space is marked.
[0074] c3. Based on the predicted vehicle occupancy information in the spatiotemporal grid map, and based on the preset conflict determination rules, identify potential conflict events between vehicles and extract the key features of each conflict event;
[0075] c4. Based on the correlation conditions of conflict events, analyze the spatiotemporal relationship between conflict events, and dynamically aggregate multiple conflict events that are closely coupled in spatiotemporal and may jointly constitute a local high-conflict-risk area into a conflict domain.
[0076] c5. For each conflict domain, analyze based on the corresponding set of conflict events, determine and output the spatiotemporal boundary of each conflict domain and the specific set of conflict events it contains.
[0077] In the above technical solution, step c1 includes:
[0078] c11, Data Extraction
[0079] The extracted data includes:
[0080] (1) Historical trajectory data of vehicles within the range of unsignalized intersections, including timestamps, longitude, latitude, speed, lateral acceleration, longitudinal acceleration, and heading angle;
[0081] (2) High-precision real-time trajectory data of vehicles within the unsignalized intersection area collected at 10Hz, including timestamp, longitude, latitude, speed, lateral acceleration, longitudinal acceleration, and heading angle;
[0082] (3) Road structure information for unsignalized intersections, including lane boundaries, stop lines, etc.;
[0083] (4) Environmental information data, including weather, road surface friction coefficient, etc.
[0084] c12, Raster Map Construction
[0085] First, a spatiotemporal cube is created using a multidimensional raster layer to initialize the three-dimensional raster data structure. Its dimensions include the X-axis (number of spatial columns), Y-axis (number of spatial rows), and T-axis (number of time segments). Each raster cell is identified as (X, Y, T), corresponding to its spatial and temporal location information.
[0086] After establishing the spatiotemporal grid structure, the spatiotemporal information of the intersection needs to be discretized. Based on road structure information, the spatial region of the intersection is divided into uniform square grids (0.5m × 0.5m). Furthermore, the observation period is divided into equal-length continuous time segments (0.1s). After spatiotemporal discretization, the timestamp and coordinate information of vehicle trajectories are mapped to the corresponding time segments and spatial grids. After information mapping, vehicle occupancy status, motion characteristics, and other attributes are updated to the corresponding spatiotemporal grid.
[0087] Step c2 includes:
[0088] c21. Dataset Construction
[0089] Load the generated spatiotemporal raster map, where each raster cell contains information such as vehicle occupancy status, velocity vector, and heading angle. Fuse the environmental information with the raster cell information to construct an enhanced dataset.
[0090] c22, Model Prediction
[0091] First, the spatiotemporal raster diagram is transformed into a spatiotemporal diagram structure, which can be specifically represented as:
[0092]
[0093] Where V represents the set of vertices (grid cells), Let represent the edge set at time t. Let represent the adjacency matrix at time t.
[0094] The adjacency matrix is constructed based on distance and calculated using a Gaussian radial basis function as the kernel function. The adjacency matrix can be specifically represented as follows:
[0095]
[0096]
[0097] in This represents the distance between node i and node j at time t. It is a predefined threshold that controls the sparsity of the adjacency matrix. It is a hyperparameter that controls the distribution.
[0098] After the adjacency matrix is constructed, the T-GCN spatiotemporal graph neural network is used for training and prediction. The temporal learning network of this model is a gated recurrent unit (GRU), and the spatial learning network is a graph convolutional network (GCN). At each time step, the GCN and GRU process the graph signal sequentially, learning the spatial and temporal correlations respectively. Figure 2 As shown, the data will pass through multiple time step units in sequence, and each unit consists of multiple steps.
[0099] First, spatial features are extracted using graph convolutional layers, which can be specifically represented as follows:
[0100]
[0101] Where X represents the node feature matrix, and A represents the adjacency matrix. This represents a trainable weight matrix. This represents a non-linear activation function.
[0102] After extracting the local spatial structure information, the update gate is entered to determine which part of the information to retain, which can be specifically represented as:
[0103]
[0104] in, This represents the graph convolution output at the current time. This indicates the hidden state at the previous moment.
[0105] The information flow then passes through a reset layer, which determines the parts of the information to be ignored. This can be specifically represented as follows:
[0106]
[0107] After determining which information to retain and which to ignore, new information can be generated, and the filtered historical information can be merged, which can be specifically represented as:
[0108]
[0109] Finally, the hidden state is updated to generate the final spatiotemporal feature representation, which can be represented as follows:
[0110]
[0111] in, This indicates the retention of historical information. This indicates newly added current information.
[0112] After traversing all time step units, the T-GCN spatiotemporal graph neural network will output the spatiotemporal probability distribution results of all vehicle trajectories.
[0113] c23, trajectory mapping
[0114] After the model outputs the results, the trajectory probability distribution is remapped onto the constructed spatiotemporal raster framework. Specifically, for k predicted trajectories, the probability distribution is first determined based on time... Perform discretization processing, and then analyze the discretized trajectory points. Mapped to the corresponding spatiotemporal raster unit This completes the filling of the spatiotemporal raster map.
[0115] Step c3 includes:
[0116] c31. Conflict Detection
[0117] Each collision event is defined as a collision between two vehicles, or a pair of vehicles. Therefore, to detect collision events, it is necessary to iterate through all pairs of vehicles. Conduct testing.
[0118] Unlike traditional methods that detect conflicts through Time-to-Cross (TTC), this invention determines conflicts by analyzing the occupancy status of spatiotemporal grid cells. Conflict detection depends on conflict points, defined as: a vehicle pair i and j have a probability of occupying the same spatiotemporal grid cell (x, y, t).
[0119] The joint occupancy probability of a vehicle within a spatiotemporal grid cell can be expressed as:
[0120]
[0121] in, This represents the probability of vehicle i occupying a spatiotemporal grid cell at (x,y,t). This represents the probability of vehicle j occupying a spatiotemporal grid cell at (x, y, t). This indicates the degree of positive correlation between the two vehicles' occupation of spatiotemporal grid cells.
[0122] If the probability of joint occupancy is greater than the set threshold (0.1), it is considered a potential conflict point.
[0123] Conflict point detection and identification requires aggregating a series of conflict points into an independent conflict event. For each vehicle pair, spatiotemporal connectivity analysis is performed to create an independent conflict point graph, which can be represented as follows:
[0124]
[0125] The collision point graph is processed using a connected component algorithm to find all collision points of the vehicle pair, where the neighborhood of the grid point (x, y, t) includes:
[0126]
[0127] in,
[0128] c32, Feature Extraction
[0129] After the conflict events are identified, feature extraction is required for each conflict event. First, the conflicting parties are determined, which can be represented as follows:
[0130]
[0131] Its time window can be specifically represented as:
[0132]
[0133]
[0134] Its spatial location can be specifically represented as:
[0135]
[0136]
[0137] Its severity can be specifically expressed as follows:
[0138]
[0139]
[0140] This led to the construction of a set of all conflict events, and the extraction of various features of each conflict event for conflict domain analysis in the case of multi-vehicle conflicts.
[0141] Step c4 includes:
[0142] c41, Conflict Association
[0143] Based on the detected set of conflict events, a correlation analysis is performed according to their characteristics, traversing all conflict event pairs to ensure that all related conflict events are extracted. For conflict events... and Correlation analysis is performed based on temporal proximity and spatial overlap conditions.
[0144] The time proximity condition is used to determine the temporal overlap of two conflicting events, and can be specifically expressed as:
[0145]
[0146]
[0147] in, This represents the time overlap threshold (taken as 1 second).
[0148] Spatial overlap condition is used to determine the spatial overlap of two conflicting events, and can be specifically expressed as:
[0149]
[0150]
[0151] in, This represents the spatial overlap threshold (taken as 5m).
[0152] The conflict event is associated if and only if both conditions are met simultaneously, and can be specifically represented as:
[0153]
[0154] c42, Connection Graph Construction
[0155] After confirming the correlation of all conflict events, in order to determine the conflict domain in the case of multiple conflicts, it is first necessary to construct an association graph. The structure of the association graph consists of nodes and edges, and its nodes are defined as each conflict event. The construction of its edges is determined by the correlation of events. If If so, then add an undirected edge between these two nodes.
[0156] A relationship diagram can be specifically represented as:
[0157]
[0158]
[0159]
[0160] After the connectivity graph is constructed, a depth-first search algorithm is used to perform connectivity component analysis, and the set of all connected components is output:
[0161]
[0162] in, Let C represent a maximal connected subgraph containing a set of interconnected conflict events. That is, each element in set C represents a set of interconnected conflict events, which we will process later and summarize as a conflict domain.
[0163] In step c5, based on the result of the previous step, each element in the set is first... Define it as a conflict domain, and then determine the various characteristics of the conflict domain based on the conflict events represented by the elements.
[0164] For a conflict domain m, based on its corresponding set of associated conflict events To conduct the analysis, we first need to determine the set of conflicting vehicles, which can be specifically represented as:
[0165]
[0166]
[0167] Next, the spatiotemporal coverage of the conflict domain is determined, and its time span can be specifically expressed as:
[0168]
[0169]
[0170] Its spatial envelope can be specifically represented as:
[0171]
[0172] like Figure 3 As shown, more detailed spatiotemporal features include the centroid location, spatial density, spatiotemporal density, and spatiotemporal heatmap, which can be represented as follows:
[0173]
[0174]
[0175]
[0176]
[0177] in, This indicates an indicator function, which takes the value 1 when a certain spatiotemporal grid coordinate falls within the range of the conflict event pair E, and takes the value 0 otherwise.
[0178] Finally, the conflict coupling strength needs to be determined, and its indicators include average conflict strength, maximum cascading conflict, and conflict strength index, which can be expressed as follows:
[0179]
[0180]
[0181]
[0182] in, This represents a randomly selected path (a series of connected conflict events) in the conflict event association graph, where |P| represents the length of path P (the number of conflict events contained in the path).
[0183] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0184] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0185] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.
[0186] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0187] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.
Claims
1. A method for collision domain detection in the case of multi-vehicle collisions at an unsignalized intersection, characterized in that, Includes the following steps: c1. Acquire historical trajectory data, real-time trajectory data, road structure information, and environmental information within unsignalized intersections. Discretize the spatial region of the intersection into a regular grid and construct a three-dimensional spatiotemporal raster map by combining it with the time dimension. Each grid cell includes vehicle attributes at the corresponding spatial location within the corresponding time segment. c2. Based on the spatiotemporal grid map, the grid cell information and environmental information are fused to construct an enhanced dataset. Based on the enhanced dataset, a pre-trained T-GCN spatiotemporal graph neural network model is used to predict the probability distribution of the driving trajectory of all vehicles in the future period. The predicted trajectory is mapped back to the spatiotemporal grid map, and the probabilistic occupancy area of each vehicle in the future spatiotemporal space is marked. c3. Based on the predicted vehicle occupancy information in the spatiotemporal grid map, identify potential conflict events between vehicles according to the preset conflict determination rules; c4. Based on the correlation conditions of conflict events, analyze the spatiotemporal relationship between conflict events, identify conflict events that are coupled in spatiotemporal space, and aggregate the interrelated conflict events to form a conflict domain; c5. For each conflict domain, analyze the corresponding set of conflict events to determine the corresponding conflict characteristics.
2. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 1, characterized in that, The historical and real-time trajectory data include timestamps, longitude, latitude, speed, lateral acceleration, longitudinal acceleration, and heading angle; the road structure information includes lane boundaries and stop lines; and the environmental information includes weather and road surface friction coefficient.
3. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 1, characterized in that, The steps for constructing the three-dimensional spatiotemporal raster map include: A spatiotemporal cube is constructed using a multidimensional raster layer to initialize a three-dimensional raster data structure. The dimensions include the number of spatial columns (X-axis), the number of spatial rows (Y-axis), and the number of time segments (T-axis). The spatial area of the intersection is divided into a uniform grid, and the observation period is divided into continuous time segments; The timestamp and coordinate information of the vehicle trajectory are mapped to the corresponding time segments and spatial grids, and the vehicle attributes are updated to the corresponding spatiotemporal grid cells.
4. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 1, characterized in that, Based on the augmented dataset, a pre-trained T-GCN spatiotemporal graph neural network model is used to predict the probability distribution of the driving trajectories of all vehicles over a future period. Specific steps include: The spatiotemporal raster graph is transformed into a spatiotemporal graph structure, which includes a vertex set, an edge set, and an adjacency matrix. Each vertex corresponds one-to-one with a raster cell, and the vertex features are defined by the enhanced feature data of the corresponding raster cell. The adjacency matrix is constructed using Gaussian radial basis functions based on the distance between the raster cells corresponding to the vertices. The spatiotemporal graph structure is input into a pre-trained T-GCN spatiotemporal graph neural network model. Spatial features are aggregated through graph convolutional layers, and temporal dependencies are learned through gated recurrent units to output hidden states that fuse spatiotemporal features. Based on the hidden states at the final moment, the probability distribution of the driving trajectories of all vehicles in the future is generated.
5. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 1, characterized in that, The steps for identifying potential conflict events between vehicles and extracting key features for each conflict event include: Traverse all vehicle pairs, calculate the joint occupancy probability of each vehicle pair within a spatiotemporal grid cell, and identify spatiotemporal grid cells whose joint occupancy probability exceeds a preset threshold as a potential conflict point of the vehicle pair. For each vehicle pair, a conflict point graph is constructed based on all its potential conflict points; spatiotemporal connectivity analysis is performed on the conflict point graph, and all potential conflict points belonging to the same connectivity are aggregated into an independent conflict event. For each independent conflict event, key features are extracted; the key features include at least one or more of the following: the vehicle pair involved in the conflict, the time window of the conflict, the spatial location of the conflict, and the severity of the conflict.
6. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 1, characterized in that, The association conditions for conflict events include temporal proximity and spatial overlap. Two conflict events are determined to be associated if and only if both temporal proximity and spatial overlap conditions are met simultaneously.
7. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 6, characterized in that, The specific steps for identifying a set of conflicting events that are coupled in space and time include: Construct an association graph with each conflict event as a node. If two conflict events are associated, add an undirected edge between the corresponding nodes. A depth-first search algorithm is used to perform connectivity component analysis on the constructed association graph, and each maximal connected subgraph obtained from the analysis is defined as a conflict domain.
8. The collision domain detection method for multi-vehicle collisions at an unsignalized intersection according to claim 1, characterized in that, The conflict characteristics include at least one of conflict vehicle set, spatiotemporal coverage, spatiotemporal features, and conflict coupling strength; the spatiotemporal features include at least one of centroid location, spatial density, spatiotemporal density, and spatiotemporal heat map; the conflict coupling strength includes at least one of average conflict strength, maximum cascading conflict, and conflict strength index.
9. A collision domain detection system for multi-vehicle collisions at an unsignalized intersection, characterized in that, To implement the method of any one of claims 1-8, comprising: Data extraction module: used to acquire historical trajectory data, real-time trajectory data, road structure information and environmental information within unsignalized intersections, discretize the spatial area of the intersection into a regular grid, and construct a three-dimensional spatiotemporal raster map by combining the time dimension; where each raster cell includes the vehicle attributes of the corresponding spatial location within the corresponding time segment; The trajectory prediction module is used to construct an enhanced dataset by fusing grid cell information and environmental information based on the spatiotemporal grid map. Based on the enhanced dataset, it uses a pre-trained T-GCN spatiotemporal graph neural network model to predict the probability distribution of the driving trajectory of all vehicles in the future period of time, maps the predicted trajectory back to the spatiotemporal grid map, and marks the probabilistic occupancy area of each vehicle in the future spatiotemporal space. Conflict identification module: used to identify potential conflict events between vehicles based on the predicted vehicle occupancy information in the spatiotemporal grid map and a preset conflict determination rule; Conflict Correlation Module: Based on the correlation conditions of conflict events, it analyzes the spatiotemporal relationships between conflict events, identifies conflict events that are coupled in spatiotemporal space, and aggregates the interrelated conflict events to form a conflict domain; Feature Analysis Module: Used to analyze each conflict domain based on the corresponding set of conflict events to determine the corresponding conflict features.
10. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the collision domain detection method for multi-vehicle collisions at an unsignalized intersection as described in any one of claims 1-8.