Traffic conflict identification method and system for road interleaving area

Abstract

The invention discloses a method and a system for identifying traffic conflicts in a road interleaving area, particularly relates to the field of traffic management, is used for solving the problems of low identification precision and early warning delay of the traffic conflicts in the interleaving area, and is used for accurately capturing risk events and filtering non-intentional behavior interference based on a composite trigger mechanism of turn light activation and transverse displacement rate; a behavior feature vector is matched with a conflict type template library to realize self-adaptive interception of an asymmetric time window, and risk evolution cycles of scenes such as sharp cut-in and slow lane change are completely covered; introducing a dynamic window expansion mechanism driven by a conflict energy gradient to ensure the whole-process capture of a risk accumulation stage; multi-scale spatio-temporal features are extracted in combination with a convolutional network to generate a thermodynamic diagram, and refined positioning of risk spatio-temporal distribution is achieved; an early warning signal grading mechanism and a closed loop feedback optimization system are established, the system adaptability is continuously improved, real-time and accurate decision support is provided for an intelligent traffic management system, the secondary accident risk is effectively restrained, and the road network traffic efficiency is optimized.

Description

Technical Field

[0001] The present invention relates to the field of traffic management, and more particularly, to a method and system for traffic conflict identification in road weaving areas. Background Art

[0002] In modern traffic systems, road weaving areas, as complex regions where traffic flows converge, diverge, and change lanes, frequently witness vehicle interactions, which are prone to triggering traffic conflict events. Especially in scenarios such as highway ramps, roundabout entrances and exits, and urban intersections, vehicles need to complete lane-changing, merging, or diverging operations within a short period. Drivers face high-frequency decision-making and coordination pressures, traffic flow changes dynamically, and the spatio-temporal distribution of risk signals exhibits asymmetry and uncertainty. Traditional traffic safety monitoring technologies mostly rely on fixed sensors or manual inspections, making it difficult to capture the microscopic interaction behaviors between vehicles in real time, and the response to the dynamic characteristics of the risk evolution process lags behind, limiting the effectiveness and timeliness of traffic safety management.

[0003] However, existing traffic conflict identification methods in road weaving areas generally use fixed time windows or static thresholds for data interception and analysis, ignoring the dynamic characteristics and forward long-tail features of the conflict event evolution chain, resulting in one-sided processing of the spatio-temporal distribution of risk signals. When analyzing triggering events, traditional methods rely on symmetric data interception at a single trigger point (such as the moment of collision prediction), failing to cover the hidden behaviors (such as speed games and lane position offsets) in the precursor stage of conflict events, thus limiting the accuracy and comprehensiveness of risk assessment. In addition, existing technologies mostly rely on trajectory data at a single scale in the feature extraction link, making it difficult to simultaneously represent microscopic trajectory fluctuations and macroscopic interaction trends, affecting the accuracy of conflict identification and the timeliness of early warning. The parameter settings of existing systems are mostly static, lacking an adaptive optimization mechanism and being difficult to cope with the dynamic changes of traffic scenarios, further restricting the practicality and reliability of the technology. To solve the above problems, a technical solution is provided herein. Summary of the Invention

[0004] To overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a traffic conflict recognition method and system for road weaving areas, which accurately capture risk events based on a composite trigger mechanism of turn signal activation and lateral displacement rate, filtering out interference from non-intentional behaviors; adopt a behavior feature vector to match a conflict type template library to achieve adaptive interception of an asymmetric time window, completely covering the risk evolution cycle of scenarios such as sharp cut-ins and slow lane changes; introduce a dynamic window expansion mechanism driven by conflict energy gradient to ensure the full process capture during the risk accumulation stage; combine a convolutional network to extract multi-scale spatio-temporal features to generate a heat map, realizing refined positioning of the risk spatio-temporal distribution; establish an early warning signal grading mechanism and a closed-loop feedback optimization system, continuously improving the system adaptability, providing real-time and accurate decision-making support for intelligent traffic management systems, effectively curbing the risk of secondary accidents and optimizing the road network traffic efficiency, so as to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A traffic conflict recognition method for road weaving areas, comprising the steps of:

[0007] S1: Use an in-vehicle camera to monitor the turn signal state and lane departure trajectory, calculate the lateral displacement rate and compare it with a dynamic threshold. When the turn signal is activated and the lateral displacement rate exceeds the standard, trigger conflict event analysis;

[0008] S2: Construct a vehicle feature vector, and determine the corresponding asymmetric time window template by similarity matching with a preset conflict type template library;

[0009] S3: Optimize the time window based on the conflict energy gradient. If the peak value of the conflict energy gradient within the window does not reach the preset threshold, gradually expand the window until the peak value reaches the threshold or meets the termination condition, and determine the time window containing the complete risk evolution chain;

[0010] S4: Collect vehicle micro-trajectory data within the optimized time window, use a pyramid convolutional network to extract spatio-temporal features, fuse them to generate a spatio-temporal conflict heat map and mark high-risk spatio-temporal coordinates;

[0011] S5: Generate a graded early warning signal according to the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically update the parameters of the asymmetric time window template in the conflict type template library through the feedback of the early warning accuracy.

[0012] In a preferred embodiment, step S1 includes the following content:

[0013] The turn signal status and lane departure trajectory of the vehicle are monitored in real time through an in-vehicle camera, the lateral displacement rate of the vehicle is calculated, and a dynamic threshold is generated based on the lane width, the vehicle's own speed, and the environmental complexity factor. When the turn signal status is activated and the lateral displacement rate exceeds the dynamic threshold, the conflict event analysis process is triggered.

[0014] In a preferred embodiment, step S1 further includes the following:

[0015] The in-vehicle camera captures the flashing situation of the turn signal to determine whether the turn signal status is activated or not. At the same time, through image processing technology, a sequence of lateral distance data of the vehicle relative to the center line of the lane is obtained, and then the lateral displacement rate is calculated. The dynamic threshold is obtained by dividing the lane width by the reference lane change time, multiplying by the ratio of the vehicle's own speed to the reference speed, and the environmental complexity factor. The trigger condition is that the turn signal status is activated and the lateral displacement rate is greater than the dynamic threshold. After triggering, the trigger time point is recorded.

[0016] In a preferred embodiment, step S2 includes the following:

[0017] A feature vector of the vehicle is constructed, where the feature vector includes the turn signal activation duration, the lateral acceleration change rate, and the lane departure angle. Among them, the turn signal activation duration is defined as the time difference from the start time of turn signal activation to the trigger time; the lateral acceleration change rate is defined as the difference between the lateral acceleration at the trigger time and the lateral acceleration of the previous frame divided by the inter-frame time difference; the lane departure angle is defined as the arctangent function value of the ratio of the lateral displacement rate to the longitudinal speed at the trigger time.

[0018] In a preferred embodiment, step S2 further includes the following:

[0019] After construction, the feature vector is matched with a preset conflict type template library, which includes three types: sharp cut-in, slow lane change, and continuous lane change. Each type corresponds to a center value of the feature vector; the matching process determines the similarity by calculating the inverse form of the Euclidean distance between the feature vector and the center value of each type of feature vector, and selects the conflict type with the highest matching degree according to the similarity value, and determines the corresponding asymmetric time window template.

[0020] In a preferred embodiment, step S3 includes the following:

[0021] Within the determined asymmetric time window template, the time series peak value of the conflict energy gradient is calculated. If the peak value does not reach the preset threshold, the window boundary is gradually expanded, that is, a fixed duration is extended each time, until the peak value of the conflict energy gradient within the window reaches the preset threshold or the total window duration reaches the preset maximum allowable duration, and finally the optimized time window is determined.

[0022] In a preferred embodiment, step S3 further includes the following content:

[0023] Among them, the conflict energy gradient is defined as the difference in the conflict risk index between adjacent time points divided by the inter-frame time difference. The conflict risk index is comprehensively calculated based on the relative distance, relative speed, and relative acceleration between the vehicle and neighboring vehicles. The window expansion preferentially expands forward to cover earlier risk accumulation stages. The expanded window is used to recalculate the peak value of the conflict energy gradient until the termination condition is met.

[0024] In a preferred embodiment, step S4 includes the following content:

[0025] Collect the microscopic trajectory data of the vehicle in the optimized time window. Use a pyramid convolutional network to extract the spatio-temporal features of the microscopic trajectory data. Capture short-term spatio-temporal features, medium-term spatio-temporal features, and long-term spatio-temporal features through multi-scale convolutional layers respectively. After aggregation by the pooling layer, fuse them to generate a comprehensive spatio-temporal feature map. Map the comprehensive spatio-temporal feature map to a risk intensity value through a fully connected layer, and generate a spatio-temporal conflict heat map through three-dimensional linear interpolation and normalization processing. In the spatio-temporal conflict heat map, mark the spatio-temporal points with risk intensity values exceeding the preset threshold as high-risk spatio-temporal coordinates.

[0026] In a preferred embodiment, step S5 includes the following content:

[0027] Obtain the spatio-temporal conflict heat map and high-risk spatio-temporal coordinates. Calculate the peak intensity of the spatio-temporal conflict heat map and the distribution density of the high-risk spatio-temporal coordinates. Generate a graded warning signal based on the peak intensity and distribution density. Record the graded warning signal and its corresponding conflict type in the conflict record database. Regularly calculate the warning accuracy rate of each conflict type. For conflict types with warning accuracy rates lower than the preset threshold, adjust by increasing the forward window duration or the backward window duration. The adjustment amplitude is a preset fixed value or dynamically determined according to the difference between the warning accuracy rate and the preset accuracy rate threshold. Update the adjusted asymmetric time window template parameters to the conflict type template library.

[0028] For a traffic conflict recognition system in a road weaving area, it includes: a trigger analysis module, a window matching module, a window optimization module, a feature extraction module, and a warning feedback module;

[0029] Trigger analysis module: Use an in-vehicle camera to monitor the turn signal status and lane departure trajectory, calculate the lateral displacement rate and compare it with a dynamic threshold. When the turn signal is activated and the lateral displacement rate exceeds the standard, trigger the conflict event analysis;

[0030] Window matching module: Construct a vehicle feature vector, and determine the corresponding asymmetric time window template through similarity matching with the preset conflict type template library;

[0031] Window Optimization Module: Optimize the time window based on the conflict energy gradient. If the peak value of the conflict energy gradient within the window does not reach the preset threshold, gradually expand the window until the peak value reaches the threshold or meets the termination condition, and determine the time window containing the complete risk evolution chain;

[0032] Feature Extraction Module: Collect vehicle microscopic trajectory data within the optimized time window, extract spatio-temporal features using a pyramid convolutional network, fuse and generate a spatio-temporal conflict heat map, and mark high-risk spatio-temporal coordinates

[0033] Early Warning Feedback Module: Generate a hierarchical early warning signal based on the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically update the asymmetric time window template parameters in the conflict type template library through the feedback of the early warning accuracy rate.

[0034] The technical effects and advantages of the present invention for the traffic conflict recognition method and system in the road weaving area:

[0035] The present invention significantly improves the traffic conflict recognition efficiency in the road weaving area through multi-dimensional technology collaboration: accurately captures risk events based on the composite trigger mechanism of turn signal activation and lateral displacement rate, effectively filters out interference from non-intentional behaviors; uses behavior feature vector matching with the conflict type template library to achieve adaptive interception of asymmetric time windows, completely covering the risk evolution cycles of scenarios such as sudden lane cuts and gradual lane changes; introduces a dynamic window expansion mechanism driven by the conflict energy gradient to ensure the full capture of the entire risk accumulation stage; combines a pyramid convolutional network to extract multi-scale spatio-temporal features to generate a high-precision heat map, realizing fine positioning of the risk spatio-temporal distribution; establishes a warning signal grading mechanism and a closed-loop feedback optimization system, dynamically calibrates the window template parameters through historical data, and continuously improves the system adaptability. Provides real-time and accurate decision support for the intelligent traffic management system, effectively curbs the risk of secondary accidents, and optimizes the road network traffic efficiency. Brief Description of the Drawings

[0036] Figure 1 It is a schematic flow chart of the traffic conflict recognition method in the road weaving area for the present invention.

[0037] Figure 2 It is a schematic structural diagram of the traffic conflict recognition system in the road weaving area for the present invention. Detailed Embodiments

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0039] Embodiment 1: Figure 1 The traffic conflict recognition method for road weaving areas of the present invention is provided, including:

[0040] S1: Use an in-vehicle camera to monitor the turn signal status and lane departure trajectory, calculate the lateral displacement rate and compare it with a dynamic threshold. When the turn signal is activated and the lateral displacement rate exceeds the standard, trigger the conflict event analysis.

[0041] S2: Construct a vehicle feature vector, and determine the corresponding asymmetric time window template through similarity matching with a preset conflict type template library.

[0042] S3: Optimize the time window based on the conflict energy gradient. If the peak value of the conflict energy gradient within the window does not reach the preset threshold, gradually expand the window until the peak value reaches the threshold or meets the termination condition, and determine the time window containing the complete risk evolution chain.

[0043] S4: Collect vehicle micro-trajectory data within the optimized time window, extract spatio-temporal features using a pyramid convolutional network, fuse them to generate a spatio-temporal conflict heat map and mark high-risk spatio-temporal coordinates.

[0044] S5: Generate a hierarchical warning signal according to the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically update the parameters of the asymmetric time window template in the conflict type template library through the feedback of warning accuracy.

[0045] In the modern urban traffic system, road weaving areas (such as highway ramps, roundabout entrances and exits, urban road intersections, etc.) have become key areas with frequent traffic accidents due to their complex traffic flow directions and dynamic vehicle interaction behaviors. Within these areas, vehicles need to frequently perform lane-changing, merging or diverging operations, and drivers must make decisions and coordinate driving with other vehicles within a short time. Due to the density of traffic flow and the unpredictability of vehicle behaviors, traditional traffic safety analysis methods have significant limitations in capturing the dynamic evolution process of conflict risks. In particular, traditional methods mostly rely on fixed time windows or static thresholds, and it is difficult to adapt to the implicit behaviors (such as vehicle speed adjustment or lane position offset) in the precursor stage of conflict events, resulting in insufficient spatio-temporal accuracy of risk warnings and prone to lags or misjudgments. In response to the traffic conflict recognition requirements in road weaving areas, the present invention proposes a dynamic analysis method based on visual recognition and intelligent algorithms, aiming to accurately capture the whole process of conflict events by real-time monitoring of vehicle behaviors and adaptively adjusting analysis parameters, so as to provide technical support for improving traffic safety.

[0046] The following is the specific processing technical logic of step S1, focusing on monitoring vehicle behaviors through an in-vehicle camera and triggering the conflict event analysis process to ensure clear connection with subsequent steps and consistent parameter definitions.

[0047] Step S1 includes the following content:

[0048] Obtain the turn signal status and lane departure trajectory of the vehicle in real time through an in-vehicle camera. The monitoring of the turn signal status relies on the in-vehicle camera to capture the blinking of the turn signal, and uses image recognition technology to determine whether the turn signal is in an activated state. If the turn signal is detected to be blinking, record the turn signal status as activated; if no blinking is detected, record the turn signal status as not activated. The monitoring of the lane departure trajectory is achieved through the image processing technology of the in-vehicle camera. Specifically, it analyzes the position offset of the vehicle relative to the lane center line to generate a sequence of lateral distance data that changes over time. Each item in the sequence of lateral distance data represents the offset distance of the vehicle from the lane center line at a certain moment, where a positive value indicates that the vehicle is biased to the right, and a negative value indicates that the vehicle is biased to the left.

[0049] Monitoring the turn signal status can directly reflect the driver's intention to change lanes or turn. As a clear signal input, it can avoid misjudging an unintentional slight offset as a lane change behavior, thus improving the accuracy of judgment.

[0050] The calculation logic of the lateral displacement rate is based on the sequence of lateral distance data of the lane departure trajectory and is used to quantify the lateral movement speed of the vehicle per unit time. The specific calculation method is as follows: Select two consecutive time points, extract the corresponding lateral distance values at these two time points from the sequence of lateral distance data, calculate the difference between the two, and then divide this difference by the time interval between the two time points to obtain the lateral displacement rate. The time interval is determined by the inter-frame time difference of data acquisition. For example, if the in-vehicle camera acquires data at a frequency of 30 frames per second, the time interval is the time required to acquire one frame. The unit of the lateral displacement rate is meters per second, and the magnitude of its value reflects the urgency of the vehicle's lane change behavior.

[0051] The calculation of the lateral displacement rate can dynamically characterize the urgency of the vehicle's lane change behavior. Compared with relying only on the static lateral distance, it can more accurately capture high-risk sudden cut-in behaviors.

[0052] The calculation logic of the dynamic threshold generates an adaptively adjusted reference value by comprehensively considering the lane width, the vehicle's own speed, and the environmental complexity factor, which is used to determine whether the lateral displacement rate is abnormal. The specific calculation method is as follows: First, take the lane width as the basic parameter, divide the lane width by a fixed reference lane-changing time, and the reference lane-changing time is set to the typical time value required for normal lane-changing; then, compare the vehicle's own speed with a fixed reference speed and calculate the ratio of the two; next, multiply the result of the previous division by the ratio of the vehicle's own speed to the reference speed; finally, multiply this result by an environmental complexity factor to obtain the dynamic threshold. The value range of the environmental complexity factor is adjusted according to the complexity of the traffic environment. For example, a higher value is taken in areas with dense traffic flow or complex road types, and a lower value is taken in simple environments. For example, it is calculated in the following way:

[0053] The dynamic threshold comprehensively considers the lane width, the vehicle's own speed, and the environmental complexity. The formula is as follows:

[0054]

[0055] Where:

[0056] W lane : The current lane width, in meters, reflecting the physical constraint of the lane-changing behavior.

[0057] T ref : The reference lane-changing time, fixed at 3 seconds, representing the time required for normal lane-changing.

[0058] V ego : The vehicle's own speed, in meters per second, affecting the dynamic characteristics of lane-changing.

[0059] V base : The reference speed, fixed at 10 m / s (about 36 km / h), for normalization.

[0060] K env : The environmental complexity factor, with a value range of [1, 1.5], adjusted according to the traffic flow density or road type (for example, a higher value is taken in the weaving area).

[0061] The introduction of the dynamic threshold avoids the defect that the traditional fixed threshold cannot adapt to different traffic scenarios. By combining the lane width, the vehicle's own speed, and the environmental complexity factor, the judgment standard can be adaptively adjusted according to the actual driving conditions. This adaptability ensures the applicability of the system in wide lanes, high-speed driving, or complex traffic environments, and improves the accuracy and flexibility of abnormal behavior recognition.

[0062] The judgment logic of the triggering condition is based on the comprehensive analysis of the turn signal state and the lateral displacement rate, and is used to determine whether to start the conflict event analysis process. The specific judgment process is as follows: First, check whether the turn signal state is activated. If the turn signal state is not activated, it is determined that the condition is not met; if the turn signal state is activated, further compare the lateral displacement rate with the dynamic threshold. If the lateral displacement rate exceeds the dynamic threshold, it is determined that the triggering condition is met, record the current moment as the triggering time point, and start the subsequent analysis process; if the lateral displacement rate does not exceed the dynamic threshold, it is determined that the condition is not met and no triggering occurs.

[0063] The dual judgment of the triggering condition combines the driver's intention and the dynamic characteristics of the lane change behavior, ensuring that the analysis is triggered only when the driver clearly expresses the intention to change lanes and the lane change behavior is extremely rapid. This mechanism can effectively filter out unintentional or normal lateral movements, reduce false triggering situations, and improve the accuracy and computational efficiency of the system.

[0064] The execution of step S1 requires meeting the preconditions, that is, the vehicle has entered the road weaving area, the on-vehicle camera has been initialized and can operate normally, continuously providing image data of the turn signal state and the lane departure trajectory. When the triggering condition is met, the processing result of step S1 is to record the triggering time point and the relevant turn signal state, lateral displacement rate, and lane departure trajectory data, which will be passed as output to the subsequent steps for further constructing the vehicle feature vector and matching the conflict type template.

[0065] The specific processing technical logic of step S1 monitors the turn signal state and the lane departure trajectory through the on-vehicle camera, calculates the lateral displacement rate and compares it with the dynamic threshold generated according to the lane width, the vehicle's own speed, and the environmental complexity factor. When the turn signal state is activated and the lateral displacement rate exceeds the dynamic threshold, the conflict event analysis is triggered. This technical logic provides an accurate triggering mechanism for traffic conflict recognition in the road weaving area. Its adaptability and dual-condition design significantly improve the accuracy, flexibility, and efficiency of the system. The output triggering time point and relevant data lay the foundation for subsequent analysis, ensuring that the entire process of the conflict event can be effectively captured and processed.

[0066] In the traffic conflict recognition in the road weaving area, step S1 monitors the turn signal state and the lane departure trajectory of the vehicle in real time through the on-vehicle camera. When the turn signal is activated and the vehicle's lateral displacement rate exceeds the dynamic threshold, the conflict event analysis process is triggered, and the triggering moment and relevant data (starting moment of turn signal activation, lateral displacement rate, lane departure trajectory) are recorded. Step S2 constructs the vehicle feature vector and matches the conflict type template library based on the data provided by S1 to determine the appropriate asymmetric time window template, providing an accurate data basis for the window optimization in step S3.

[0067] Step S2 includes the following:

[0068] Starting from the data obtained in step S1, the processing logic for constructing the feature vector calculates the turn signal activation duration, the lateral acceleration change rate, and the lane departure angle, and generates a vehicle behavior feature vector.

[0069] Turn signal activation duration: Using the start time of turn signal activation monitored by the on-vehicle camera and the trigger time recorded in step S1, calculate the time difference between the two. Specifically, directly subtract the time interval from the start time of turn signal activation to the trigger time, and the resulting value is the turn signal activation duration, which is used to characterize the urgency of the driver's lane change decision.

[0070] Lateral acceleration change rate: First, calculate the lateral acceleration based on the lateral displacement rate provided in step S1. The method is to subtract the lateral displacement rate of the current frame from the lateral displacement rate of the previous frame, and then divide by the inter-frame time difference of data acquisition to obtain the lateral acceleration. Then, take the difference between the lateral acceleration at the trigger time and the lateral acceleration of the previous frame, and divide by the inter-frame time difference to obtain the lateral acceleration change rate, which is used to characterize the rapidity of the lane change behavior.

[0071] For example, the lateral acceleration change rate A lat_rate :

[0072] Is defined as the change rate of the vehicle's lateral acceleration per unit time, characterizing the rapidity of the lane change behavior.

[0073]

[0074] A lat (t trigger ): The lateral acceleration at the trigger time, calculated by differentiating the lateral displacement rate V lat (t trigger ).

[0075] A lat (t trigger - Δt): The lateral acceleration of the previous frame before the trigger time.

[0076] Δt: The fixed inter-frame time difference of data acquisition, determined by the sampling frequency of the on-vehicle camera.

[0077] V lat (t trigger ): The lateral displacement rate at the trigger time in step S1.

[0078] Lane departure angle: Using the ratio of the lateral displacement rate to the longitudinal speed at the triggering moment, calculate the value of the arctangent function and convert it to degrees. Specifically, divide the lateral displacement rate by the longitudinal speed, calculate the result of the arctangent function, and then multiply the result by 180 divided by pi to obtain the angle value, which is used to reflect the included angle between the vehicle's driving direction and the lane line direction.

[0079] For example, the lane departure angle θ dev :

[0080] is defined as the included angle between the vehicle's driving direction and the lane line direction, reflecting the degree of deviation of the vehicle from the lane line.

[0081]

[0082] V lat (t trigger ): The lateral displacement rate at the triggering moment.

[0083] V long (t trigger ): The longitudinal speed at the triggering moment, that is, the vehicle's own speed, recorded by on-vehicle sensors.

[0084] Finally, combine the turn signal activation duration, the lateral acceleration change rate, and the lane departure angle into a vehicle behavior feature vector.

[0085] By constructing a vehicle behavior feature vector, it is possible to comprehensively quantify the decision-making urgency, rapidity, and deviation degree of the vehicle's lane-changing behavior, providing a multi-dimensional behavior description.

[0086] The processing logic of the conflict type template library predefines three conflict types, namely sudden cut-in, slow lane change, and continuous lane change. Each type corresponds to a feature vector range and an asymmetric time window template. The feature vector range of the sudden cut-in type is defined as a short turn signal activation duration, a high lateral acceleration change rate, and a large lane departure angle. Its corresponding asymmetric time window template is designed with a long forward window and a short backward window. Specifically, the forward window covers a longer time period before the triggering moment, and the backward window covers a shorter time period after the triggering moment to capture the risk accumulation stage before a rapid lane change. The feature vector range of the slow lane change type is defined as a medium turn signal activation duration, a low lateral acceleration change rate, and a small lane departure angle. Its corresponding asymmetric time window template is designed with both the forward window and the backward window being medium. Specifically, the time spans of both are similar to balance the coverage of the risk evolution process. The feature vector range of the continuous lane change type is defined as a long turn signal activation duration, a fluctuating lateral acceleration change rate, and a frequently changing lane departure angle. Its corresponding asymmetric time window template is designed with both the forward window and the backward window being long. Specifically, both cover a long time period to include the risk evolution process of multiple lane change behaviors.

[0087] Through the predefined conflict type template library, it is possible to quickly classify according to the characteristics of the vehicle behavior feature vector and select the matching asymmetric time window template.

[0088] The processing logic of feature vector matching is quantified by calculating the similarity between the vehicle behavior feature vector and the central value of each type of feature vector in the conflict type template library in the form of the inverse function of the Euclidean distance. The specific method is as follows: First, calculate the differences in the turn signal activation duration, the lateral acceleration change rate, and the lane departure angle between the vehicle behavior feature vector and the central value of each conflict type feature vector; then, calculate the squares of these three differences respectively, and then add the three squared values to obtain the sum of squares; next, take the square root of the sum of squares to obtain the Euclidean distance; finally, take the reciprocal after adding 1 to the Euclidean distance to obtain the similarity value, and the range of the similarity value is between 0 and 1, and the larger the value, the higher the matching degree. After completing the above calculations, select the conflict type with the highest similarity value and use the corresponding asymmetric time window template of this conflict type as the basis for subsequent analysis.

[0089] The logic of the overall processing flow is executed in the following order: First, obtain data such as the trigger moment, the starting moment of turn signal activation, the lateral displacement rate, and the longitudinal speed from step S1; then, calculate the turn signal activation duration using the starting moment of turn signal activation and the trigger moment; next, calculate the lateral acceleration change rate based on the lateral displacement rate and calculate the lane departure angle in combination with the longitudinal speed; then, combine the turn signal activation duration, the lateral acceleration change rate, and the lane departure angle into a vehicle behavior feature vector; then, calculate the similarity between the vehicle behavior feature vector and the central value of each type of feature vector in the conflict type template library; finally, select the conflict type with the highest matching degree according to the similarity value and determine its corresponding asymmetric time window template.

[0090] The overall processing flow organically integrates data acquisition, feature calculation, and template matching to form an adaptive asymmetric time window selection mechanism. This method can quickly respond and provide a suitable time window in different conflict scenarios, providing a solid foundation for subsequent risk evolution analysis.

[0091] The execution of step S2 depends on the data such as the trigger moment, the starting moment of turn signal activation, the lateral displacement rate, and the longitudinal speed provided by step S1 to ensure the real-time and accurate construction of the vehicle behavior feature vector. After completing the feature vector matching, the determined asymmetric time window template will be passed to step S3 for optimizing and expanding the time window to support a more comprehensive risk assessment process.

[0092] Step S2 realizes the adaptive selection of the asymmetric time window template by constructing a vehicle behavior feature vector and performing similarity matching with a conflict type template library. It can accurately adapt to the requirements of different conflict scenarios, providing a flexible and accurate analysis basis for traffic conflict recognition in road weaving areas, and significantly improving the pertinence and comprehensiveness of risk assessment.

[0093] Based on the vehicle behavior feature vector, Step S2 matches a preset conflict type template library to determine an initial asymmetric time window template, providing a preliminary time range for risk evolution analysis. However, due to the complexity of traffic scenarios and the dynamic changes in vehicle behavior, the asymmetric time window template determined in Step S2 may not fully cover the entire process of the conflict risk evolution chain, especially when the peak of the risk signal is not fully manifested or the evolution chain of the conflict event is relatively long. Therefore, Step S3 introduces a window optimization and extension mechanism based on the conflict energy gradient. By analyzing the time series peak of the conflict energy gradient within the window in real time and dynamically adjusting the window boundary, it ensures that the finally determined time window can completely contain the key stages of the risk evolution chain.

[0094] Based on the asymmetric time window template determined in Step S2, Step S3 judges whether the risk evolution is fully manifested by calculating the time series peak of the conflict energy gradient within the window and comparing it with a preset threshold. If the peak does not reach the preset threshold, the window boundary is gradually expanded until the peak reaches the threshold or meets the termination condition, and finally a time window containing the complete risk evolution chain is determined. This optimized time window will be directly passed to Step S4 for subsequent collection and feature extraction of microscopic trajectory data.

[0095] Step S3 includes the following:

[0096] The processing logic of conflict energy gradient calculation aims to dynamically evaluate the trend of traffic conflict risk evolution by quantifying the change rate of conflict risk indicators per unit time. The specific method is as follows:

[0097] First, calculate the conflict risk indicator. The conflict risk indicator is a quantity comprehensively evaluated based on the relative motion parameters of a vehicle and neighboring vehicles. The calculation process is to multiply the reciprocal of the relative distance between the vehicle and neighboring vehicles by the weighted sum of the relative speed and relative acceleration. The reciprocal of the relative distance is used to reflect the degree of proximity between vehicles, and the relative speed and relative acceleration are used to characterize the difference in the motion states between vehicles. Then, the conflict energy gradient is defined as the difference between the conflict risk indicators at adjacent time points divided by the inter-frame time difference of data acquisition. The unit of the obtained result is the change rate of risk indicators per second, which is used to characterize the dynamic change speed of risk evolution. During the calculation process, the conflict risk indicators at adjacent time points are obtained from continuously collected vehicle motion data, and the inter-frame time difference is determined by the sampling frequency of the data acquisition device.

[0098] For example, the calculation of the conflict risk indicator can be as follows:

[0099] Conflict risk indicator R conflict (t) is comprehensively evaluated based on the relative motion parameters between vehicles, and the formula is:

[0100]

[0101] d rel (t): The relative distance between the vehicle at time t and the adjacent vehicle, which is obtained by real-time monitoring of the on-vehicle camera.

[0102] v rel (t): The relative speed between the vehicle at time t and the adjacent vehicle, which is calculated by differentiating the trajectory data.

[0103] a rel (t): The relative acceleration between the vehicle at time t and the adjacent vehicle, which is calculated by the second-order differentiation of the speed data.

[0104] The calculation of the conflict energy gradient can monitor the trend of the intensification or mitigation of risk evolution in real time. Compared with relying only on static conflict risk indicators, this method can more accurately capture the key evolution stages of traffic conflict events. By dynamically quantifying the speed of risk change, it provides a reliable basis for the optimization of subsequent time windows, ensuring that the window adjustment can reflect the actual needs of risk evolution.

[0105] The processing logic of peak detection within the time window is based on the asymmetric time window template determined in step S2. Analyze the time series of the conflict energy gradient within the window, identify its maximum value and compare it with a preset threshold. The specific method is as follows:

[0106] First, obtain the asymmetric time window template provided in step S2 and determine the initial time window, which covers a specific time period before and after the triggering moment; then, within this initial time window, calculate the time series of the conflict energy gradient, which consists of the conflict energy gradient values at multiple consecutive time points, and find the maximum value of the conflict energy gradient from it; then, compare this maximum value with the preset conflict energy gradient threshold. The preset conflict energy gradient threshold is a reference value determined based on historical traffic conflict data and expert experience, and is used to judge whether the risk evolution is fully manifested within the current window.

[0107] When the maximum value of the conflict energy gradient within the time window does not reach the preset conflict energy gradient threshold, the processing logic of the window extension mechanism ensures the integrity of the traffic conflict risk evolution chain by gradually expanding the window boundary to cover a wider time period. The specific method is as follows:

[0108] When it is detected that the maximum value of the conflict energy gradient within the time window is less than the preset conflict energy gradient threshold, a window expansion operation is triggered. The time window boundary is preferentially expanded forward to cover earlier stages of traffic conflict risk accumulation. The duration of each expansion is a fixed value, for example, increasing by 0.5 seconds each time. After expansion, the conflict energy gradient time series within the new window is recalculated, and the new maximum value is found. The expansion process continues until the maximum value of the conflict energy gradient within the window reaches or exceeds the preset conflict energy gradient threshold, or the total window duration reaches the preset maximum allowable duration. The maximum allowable duration is the upper limit set by the system to prevent excessive window expansion.

[0109] The window expansion mechanism can adapt to the risk evolution duration of different traffic conflict scenarios by dynamically adjusting the window boundary, ensuring that the analysis window completely includes the key risk stage. It improves the system's ability to identify long-term traffic conflict events. At the same time, by setting the maximum allowable duration as the termination condition, it avoids waste of computing resources caused by unlimited expansion.

[0110] During the window expansion process of the processing logic for determining the final time window, when the maximum value of the conflict energy gradient reaches the preset conflict energy gradient threshold or the total window duration reaches the maximum allowable duration, the expansion stops and the optimized time window is determined. The specific method is as follows:

[0111] When the maximum value of the conflict energy gradient within the expanded time window reaches or exceeds the preset conflict energy gradient threshold, it indicates that the traffic conflict risk evolution chain has been fully manifested. At this time, the current time window is determined as the final optimized time window. If the total duration of the time window reaches the preset maximum allowable duration, but the maximum value of the conflict energy gradient still does not reach the preset conflict energy gradient threshold, the time window that reaches the maximum allowable duration is used as the final optimized time window to ensure the timeliness of the analysis process.

[0112] The determination of the final time window, through clear termination conditions, while ensuring the integrity of the traffic conflict risk evolution chain, controls the duration of the time window to ensure that the system achieves a balance between real-time performance and accuracy.

[0113] The execution of step S3 depends on the asymmetric time window template provided by step S2 as the starting point of the initial time window to ensure the pertinence of window optimization. After completing the time window optimization, the output final optimized time window is directly passed to step S4 to collect vehicle micro-trajectory data and perform spatio-temporal feature extraction, ensuring the continuity and consistency of the entire analysis process.

[0114] The specific processing technical logic of step S3, through the calculation of the conflict energy gradient and the dynamic expansion mechanism of the time window, ensures that the finally optimized time window can completely cover the key stages of the traffic conflict risk evolution chain in the road weaving area. It can adapt to the risk evolution duration of different traffic conflict scenarios, providing accurate time range support for the subsequent collection of vehicle micro-trajectory data and the extraction of spatio-temporal characteristics, and significantly improving the accuracy and comprehensiveness of traffic conflict recognition.

[0115] Step S3 optimizes and expands the time window through the conflict energy gradient, determining an optimized time window that includes the complete risk evolution chain, providing an accurate time range for subsequent feature extraction. However, optimizing only the time window is not sufficient to comprehensively characterize the spatio-temporal distribution of traffic conflict risks. It is necessary to further extract key features from the micro-trajectory data to capture the interaction behaviors between vehicles and potential conflict points. Therefore, in the optimized time window determined in step S3, step S4 collects vehicle micro-trajectory data and extracts spatio-temporal characteristics to generate a spatio-temporal conflict heat map, marking high-risk spatio-temporal coordinates, providing a visual risk assessment basis for the generation of the hierarchical warning signal in step S5.

[0116] In the optimized time window determined in step S3, step S4 collects the micro-trajectory data of vehicles, including position, speed, and acceleration, uses a pyramid convolutional network to extract the spatio-temporal characteristics of these data, fuses them to generate a spatio-temporal conflict heat map, and marks the high-risk spatio-temporal coordinates. This step provides a visual risk assessment tool for step S5, ensuring that the spatio-temporal distribution of traffic conflict risks is accurately characterized.

[0117] Step S4 includes the following:

[0118] The processing logic of micro-trajectory data collection: within the optimized time window determined in step S3, the position, speed, and acceleration data of vehicles are collected in real time through on-vehicle cameras and sensors.

[0119] The collected data is first preprocessed. The preprocessing process includes two steps: denoising and interpolation. The denoising operation uses a moving average filtering technique to smooth the data by calculating the average value of consecutive data points, eliminating the influence of sensor noise; the interpolation operation uses a linear interpolation technique to fill in the missing data points proportionally between adjacent data points in time to generate continuous time series data, ensuring the data integrity in the subsequent feature extraction process.

[0120] The process of micro-trajectory data collection can comprehensively describe the behavioral characteristics of vehicles in traffic conflict events by collecting multi-dimensional dynamic data of vehicles. The denoising and interpolation steps in the preprocessing improve the quality and continuity of the data, providing a reliable data basis for subsequent feature extraction, thus enhancing the accuracy and robustness of traffic conflict analysis.

[0121] The processing logic of the pyramid convolution network structure extracts spatio-temporal features from microscopic trajectory data by designing multi-scale convolutional layers. The network structure consists of four parts: an input layer, a convolutional layer, a pooling layer, and a feature fusion layer. The input layer receives preprocessed microscopic trajectory data, and the data is organized in the form of the number of vehicles, the number of frames within a time window, and the feature dimensions, where the feature dimensions include position, speed, and acceleration. The convolutional layer uses convolutional kernels of different sizes, such as 1x1, 3x3, and 5x5, to extract short-term, medium-term, and long-term features respectively; specifically, smaller convolutional kernels are used to capture local changes in vehicle trajectories, and larger convolutional kernels are used to capture global trends in vehicle trajectories. The pooling layer adopts the max pooling technique to aggregate features by selecting the maximum value within a local area, reducing the data dimension while retaining the main feature information. The feature fusion layer integrates the outputs of different convolutional layers through a concatenation operation to generate a comprehensive spatio-temporal feature map containing multi-scale information, which is used to reflect the interaction behavior between vehicles and potential traffic conflict points.

[0122] Through multi-scale feature extraction technology, the pyramid convolution network structure can capture both local details and overall trends of vehicle trajectories, ensuring that the spatio-temporal features of traffic conflict risks are comprehensively characterized. This network design improves the accuracy and adaptability of feature extraction, providing a solid foundation for subsequent generation of accurate spatio-temporal conflict heatmaps.

[0123] The processing logic for generating the spatio-temporal conflict heatmap is based on the spatio-temporal features extracted by the pyramid convolution network and generates a three-dimensional heatmap through three steps: risk intensity mapping, spatio-temporal interpolation, and normalization. First, the spatio-temporal feature map is processed through a fully connected layer to convert the features into risk intensity values, and the risk intensity value represents the degree of traffic conflict risk at a specific time and space position; the role of the fully connected layer is to perform weighted combination on the input features and output a value reflecting the risk magnitude. Then, three-dimensional linear interpolation is performed on the risk intensity values. By calculating intermediate values proportionally between adjacent spatio-temporal points, continuous spatio-temporal distribution data is generated to ensure the smoothness of the heatmap in the time and space dimensions. Finally, the generated heatmap data is normalized to the range of 0 to 1, and the normalization process is achieved by dividing all risk intensity values by the maximum value, which is convenient for subsequent risk threshold judgment and evaluation.

[0124] The process of generating the spatio-temporal conflict heatmap converts spatio-temporal features into a visualized risk distribution map, intuitively showing the spatio-temporal variation characteristics of traffic conflict risks. Interpolation and normalization processing ensure the continuity and comparability of the heatmap, improving the accuracy and practical application value of risk identification.

[0125] The processing logic for marking high-risk spatio-temporal coordinates is based on the spatio-temporal conflict heatmap, identifying spatio-temporal points where the risk intensity exceeds a preset threshold and recording them as high-risk spatio-temporal coordinates. The specific process:

[0126] First, a preset risk intensity threshold needs to be set, which is determined based on the statistical analysis of historical traffic conflict data and expert experience, and the unit is consistent with the risk intensity value in the heat map. Next, traverse each spatio-temporal point in the spatio-temporal conflict heat map and compare its risk intensity value with the preset threshold one by one; if the risk intensity value of a certain spatio-temporal point exceeds the preset threshold, record the coordinates of this spatio-temporal point as the high-risk spatio-temporal coordinates.

[0127] The processing logic of step S4 uses the optimized time window provided in step S3 as the starting point for data collection to ensure that the analysis process is consistent with the time range of the previous steps. After generating the spatio-temporal conflict heat map and high-risk spatio-temporal coordinates, the processing results are directly passed to step S5 for generating the graded warning signal, ensuring the seamless connection of the analysis process of the entire system.

[0128] The specific processing technical logic of step S4 is to collect vehicle micro-trajectory data within the optimized time window, extract spatio-temporal features, generate a spatio-temporal conflict heat map, and mark high-risk spatio-temporal coordinates. This method can accurately characterize the spatio-temporal distribution characteristics of traffic conflict risks in the road weaving area, providing a visual and reliable risk assessment basis for the subsequent generation of graded warning signals, thus significantly improving the accuracy and application value of traffic conflict recognition.

[0129] Marking the high-risk spatio-temporal coordinates is to provide accurate spatio-temporal positioning information for the subsequent generation of graded warning signals, ensuring that the system can accurately identify and warn the critical moments and locations of traffic conflict events. By identifying and recording the spatio-temporal points in the spatio-temporal conflict heat map whose risk intensity exceeds the preset threshold, the system transforms the abstract risk assessment into specific time and space coordinates, thus providing intuitive and operable risk prompts for drivers or traffic management systems. This marking method enhances the pertinence and timeliness of the warning signal, and at the same time provides data support for traffic safety management, helping to prevent and reduce the occurrence of traffic accidents.

[0130] Step S4 collects vehicle micro-trajectory data within the optimized time window, extracts spatio-temporal features, generates a spatio-temporal conflict heat map and high-risk spatio-temporal coordinates, providing a visual basis for risk assessment. However, relying solely on the spatio-temporal conflict heat map and marking high-risk points is not enough to achieve real-time traffic safety warning. It is necessary to further transform the risk information into graded warning signals and optimize the asymmetric time window template parameters in step S2 through a feedback mechanism to adapt to the dynamic traffic environment. Therefore, step S5 needs to clarify the processing logic based on the spatio-temporal conflict heat map and high-risk spatio-temporal coordinates, generate operable warning signals and achieve closed-loop optimization.

[0131] Step S5 obtains the spatio-temporal conflict heat map and high-risk spatio-temporal coordinates from step S4, generates a hierarchical warning signal based on the peak intensity and distribution density, and feeds back the warning result to the conflict type template library in step S2 to dynamically update the parameters of the asymmetric time window template, forming a closed-loop optimization mechanism.

[0132] Step S5 includes the following content:

[0133] The generation of the hierarchical warning signal determines the severity of the traffic conflict risk by analyzing the peak intensity of the spatio-temporal conflict heat map and the distribution density of the high-risk spatio-temporal coordinates, and generates the corresponding warning signal.

[0134] The specific method is as follows:

[0135] First, extract the peak intensity from the spatio-temporal conflict heat map generated in step S4. The peak intensity is defined as the maximum value of the risk intensity value in the spatio-temporal conflict heat map, which is used to reflect the highest point of the traffic conflict risk. Secondly, calculate the distribution density of the high-risk spatio-temporal coordinates. The distribution density is obtained by counting the number of high-risk spatio-temporal coordinates and then dividing it by the product of the spatial area of the road weaving area and the time window duration, which is used to reflect the concentration degree of high-risk points in space and time. Then, combine the peak intensity and the distribution density, and use a preset threshold matrix for hierarchical judgment. The hierarchical rule is: when the peak intensity is lower than the preset low threshold or the distribution density is lower than the preset low threshold, the warning signal is set to a low level; when the peak intensity is between the preset low threshold and high threshold and the distribution density is between the preset low threshold and high threshold, the warning signal is set to a medium level; when the peak intensity is higher than the preset high threshold and the distribution density is higher than the preset high threshold, the warning signal is set to a high level.

[0136] The process of generating the hierarchical warning signal can comprehensively evaluate the severity of traffic conflicts by comprehensively analyzing the peak intensity and distribution density, providing intuitive and operable warning information. It avoids misjudgments that may be caused by a single indicator, improves the accuracy and practicality of the warning signal, provides a basis for drivers or traffic management systems to respond to traffic risks in a timely manner, and thus reduces the possibility of traffic accidents.

[0137] The feedback mechanism and parameter update record the hierarchical warning signal and its corresponding conflict type, regularly analyze the warning accuracy rate, and adjust the parameters of the asymmetric time window template in step S2 according to the analysis results. The specific method is as follows:

[0138] First, store the generated hierarchical warning signals each time and their corresponding conflict types in the conflict record database to form historical data accumulation. Then, regularly extract data from the conflict record database and calculate the warning accuracy rate for each conflict type. The warning accuracy rate is obtained by counting the number of times the warning signal is consistent with the actual conflict event and then dividing it by the total number of warning signals generated under this conflict type. Next, for the conflict types with warning accuracy rates lower than the preset accuracy rate threshold, adjust the corresponding asymmetric time window template parameters in step S2. The adjustment method is to increase the forward window duration or the backward window duration of the asymmetric time window template. The adjustment amplitude can be a preset fixed value or dynamically determined according to the difference between the warning accuracy rate and the preset accuracy rate threshold. Finally, update the adjusted asymmetric time window template parameters to the conflict type template library in step S2 and continuously monitor the change of the subsequent warning accuracy rate.

[0139] Through the analysis of historical data and the dynamic adjustment of parameters, the feedback mechanism and the parameter update process ensure that the asymmetric time window template can adapt to the actual changes in the traffic environment. This method realizes the self-adaptive optimization of the system, improves the accuracy and stability of traffic conflict recognition, and provides reliable support for the long-term operation of traffic safety management.

[0140] The processing of the closed-loop optimization mechanism forms an adaptive optimization cycle by continuously monitoring the warning accuracy rate and dynamically adjusting the parameters of the asymmetric time window template. The specific method is as follows:

[0141] After completing the update of the asymmetric time window template parameters, continuously collect the subsequent generated hierarchical warning signals and their corresponding conflict type data, and calculate the warning accuracy rate. If the calculated warning accuracy rate is still lower than the preset accuracy rate threshold, further adjust the forward window duration or the backward window duration of the asymmetric time window template. The adjustment process is repeated until the warning accuracy rate meets the requirements of the preset accuracy rate threshold. This mechanism ensures that the system can continuously optimize its own analysis parameters to adapt to the complex changes in the traffic environment.

[0142] Through continuous monitoring and adjustment, the closed-loop optimization mechanism realizes the self-improvement of the traffic conflict recognition system, ensuring its accuracy and timeliness in different traffic scenarios. It enhances the robustness and reliability of the system and provides technical guarantee for the continuous improvement of traffic safety management.

[0143] The processing logic of step S5 uses the spatio-temporal conflict heat map and high-risk spatio-temporal coordinates provided by step S4 as input data to ensure data consistency in the analysis process with the previous steps. After generating the graded warning signals and adjusting the parameters of the asymmetric time window template, the updated parameters of the asymmetric time window template are directly fed back to the conflict type template library in step S2 for subsequent conflict type identification and analysis. This design ensures that the entire system forms a complete adaptive analysis loop after a traffic conflict event is triggered.

[0144] The specific processing technical logic of step S5 generates graded warning signals by analyzing the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically adjusts the parameters of the asymmetric time window template based on the warning accuracy rate to form a closed-loop optimization mechanism. This method realizes real-time warning and adaptive optimization of traffic conflicts in road weaving areas, improves the efficiency and accuracy of traffic safety management, and provides technical support for the intelligent development of the traffic system.

[0145] Embodiment 2: Figure 2 This invention provides a traffic conflict identification system for road weaving areas, including:

[0146] A trigger analysis module, a window matching module, a window optimization module, a feature extraction module, and a warning feedback module;

[0147] The trigger analysis module: uses an in-vehicle camera to monitor the turn signal status and lane departure trajectory, calculates the lateral displacement rate and compares it with a dynamic threshold. When the turn signal is activated and the lateral displacement rate exceeds the standard, it triggers the analysis of conflict events;

[0148] The window matching module: constructs a vehicle feature vector and determines the corresponding asymmetric time window template through similarity matching with a preset conflict type template library;

[0149] The window optimization module: optimizes the time window based on the conflict energy gradient. If the peak value of the conflict energy gradient within the window does not reach the preset threshold, the window is gradually expanded until the peak value reaches the threshold or meets the termination condition to determine the time window containing the complete risk evolution chain;

[0150] The feature extraction module: collects vehicle micro-trajectory data within the optimized time window, uses a pyramid convolutional network to extract spatio-temporal features, fuses them to generate a spatio-temporal conflict heat map and marks high-risk spatio-temporal coordinates

[0151] The warning feedback module: generates graded warning signals according to the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically updates the parameters of the asymmetric time window template in the conflict type template library through warning accuracy feedback.

[0152] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulations to get a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0153] It should be noted that the system of the present invention can be deployed on the device itself to achieve embedded applications, or can also run on a PC or other terminal with a user interface, so as to meet various hardware environments and usage requirements.

[0154] Only some exemplary embodiments of the present invention have been described above by way of illustration. Undoubtedly, for those of ordinary skill in the art, without departing from the spirit and scope of the present invention, the described embodiments can be modified in various different ways. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0155] It should be noted that in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or device including the element.

[0156] The above is only the specific implementation manner of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art within the technical scope disclosed in this application can easily think of changes or substitutions, which should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. A traffic conflict recognition method for road weaving sections, characterized in that Including the steps: S1: Monitor the turn signal status and lane departure trajectory using an in-vehicle camera, calculate the lateral displacement rate and compare it with a dynamic threshold. When the turn signal is activated and the lateral displacement rate exceeds the standard, trigger the conflict event analysis; S2: Construct a vehicle feature vector, and determine the corresponding asymmetric time window template through similarity matching with a preset conflict type template library; S3: Optimize the time window based on the conflict energy gradient. If the peak value of the conflict energy gradient within the window does not reach the preset threshold, gradually expand the window until the peak value reaches the threshold or meets the termination condition, and determine the time window containing the complete risk evolution chain; S4: Collect vehicle microscopic trajectory data within the optimized time window, extract spatio-temporal features using a pyramid convolutional network, fuse them to generate a spatio-temporal conflict heat map and mark high-risk spatio-temporal coordinates; S5: Generate a hierarchical warning signal according to the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically update the parameters of the asymmetric time window template in the conflict type template library through feedback of the warning accuracy.

2. The traffic conflict recognition method for road weaving areas according to claim 1, wherein Step S1 includes the following: Monitor the turn signal status and lane departure trajectory of the vehicle in real time through an in-vehicle camera, calculate the lateral displacement rate of the vehicle, and generate a dynamic threshold based on the lane width, the vehicle's own speed, and the environmental complexity factor. When the turn signal status is activated and the lateral displacement rate exceeds the dynamic threshold, trigger the conflict event analysis process.

3. The traffic conflict recognition method for road weaving areas according to claim 2, characterized in that Step S1 also includes the following: The in-vehicle camera captures the flashing situation of the turn signal to determine whether the turn signal status is activated or not. At the same time, through image processing technology, obtain the lateral distance data sequence of the vehicle relative to the lane center line, and then calculate the lateral displacement rate. The dynamic threshold is obtained by dividing the lane width by the reference lane change time, multiplying by the ratio of the vehicle's own speed to the reference speed, and the environmental complexity factor. The trigger condition is that the turn signal status is activated and the lateral displacement rate is greater than the dynamic threshold. After triggering, record the trigger time point.

4. The traffic conflict recognition method for road weaving areas according to claim 3, characterized in that, Step S2 includes the following: Construct a vehicle feature vector, where the feature vector includes the turn signal activation duration, the lateral acceleration change rate, and the lane departure angle. Among them, the turn signal activation duration is defined as the time difference from the start time of turn signal activation to the trigger time; the lateral acceleration change rate is defined as the difference between the lateral acceleration at the trigger time and the lateral acceleration of the previous frame divided by the inter-frame time difference; the lane departure angle is defined as the arctangent function value of the ratio of the lateral displacement rate to the longitudinal speed at the trigger time.

5. The traffic conflict recognition method for road weaving areas according to claim 4, characterized in that, Step S2 also includes the following: After construction, match the feature vector with a preset conflict type template library. The conflict type template library includes three types: sudden cut-in, slow lane change, and continuous lane change. Each type corresponds to a feature vector center value; the matching process determines the similarity by calculating the inverse form of the Euclidean distance between the feature vector and the center value of each type of feature vector, and selects the conflict type with the highest matching degree according to the similarity value to determine its corresponding asymmetric time window template.

6. The traffic conflict recognition method for road weaving areas according to claim 5, characterized in that Step S3 includes the following: Within a determined asymmetric time window template, calculate the peak value of the time series of the conflict energy gradient. If the peak value does not reach the preset threshold, gradually expand the window boundary, that is, expand a fixed duration each time, until the peak value of the conflict energy gradient within the window reaches the preset threshold or the total window duration reaches the preset maximum allowable duration, and finally determine the optimized time window.

7. The traffic conflict recognition method for road weaving areas according to claim 6, wherein Step S3 also includes the following: Among them, the conflict energy gradient is defined as the difference of the conflict risk index between adjacent time points divided by the inter-frame time difference, and the conflict risk index is comprehensively calculated based on the relative distance, relative speed and relative acceleration between the vehicle and neighboring vehicles; The window expansion preferentially expands forward to cover earlier risk accumulation stages, and the expanded window is used to recalculate the peak value of the conflict energy gradient until the termination condition is met.

8. The traffic conflict recognition method for road weaving areas according to claim 7, characterized in that Step S4 includes the following: Collect the microscopic trajectory data of the vehicle within the optimized time window, use a pyramid convolutional network to extract the spatio-temporal features of the microscopic trajectory data, capture short-term spatio-temporal features, medium-term spatio-temporal features and long-term spatio-temporal features through multi-scale convolutional layers respectively, and fuse and generate a comprehensive spatio-temporal feature map after aggregation by the pooling layer; map the comprehensive spatio-temporal feature map to a risk intensity value through a fully connected layer, and generate a spatio-temporal conflict heat map through three-dimensional linear interpolation and normalization processing; In the spatio-temporal conflict heat map, mark the spatio-temporal points with risk intensity values exceeding the preset threshold as high-risk spatio-temporal coordinates.

9. The traffic conflict recognition method for road weaving areas according to claim 8, wherein Step S5 includes the following: Obtain the spatio-temporal conflict heat map and high-risk spatio-temporal coordinates, calculate the peak intensity of the spatio-temporal conflict heat map and the distribution density of the high-risk spatio-temporal coordinates, generate a hierarchical warning signal according to the peak intensity and the distribution density, record the hierarchical warning signal and its corresponding conflict type in the conflict record database, regularly calculate the warning accuracy rate of each conflict type, for the conflict type with the warning accuracy rate lower than the preset threshold, adjust it by increasing the forward window duration or the backward window duration, and the adjustment range is a preset fixed value or dynamically determined according to the difference between the warning accuracy rate and the preset accuracy rate threshold, and update the parameters of the asymmetric time window template after adjustment to the conflict type template library.

10. A traffic conflict recognition system for a road weaving area, which is used to implement the traffic conflict recognition method for a road weaving area described in any one of claims 1-9, characterized in that, including: a trigger analysis module, a window matching module, a window optimization module, a feature extraction module and a warning feedback module; Trigger analysis module: Use an in-vehicle camera to monitor the turn signal status and lane departure trajectory, calculate the lateral displacement rate and compare it with a dynamic threshold. When the turn signal is activated and the lateral displacement rate exceeds the standard, trigger the conflict event analysis; Window matching module: Construct a vehicle feature vector, and determine the corresponding asymmetric time window template through similarity matching with the preset conflict type template library; Window optimization module: Optimize the time window based on the conflict energy gradient. If the peak value of the conflict energy gradient within the window does not reach the preset threshold, gradually expand the window until the peak value reaches the threshold or the termination condition is met, and determine the time window containing the complete risk evolution chain; Feature extraction module: Collect the microscopic trajectory data of the vehicle within the optimized time window, use a pyramid convolutional network to extract spatio-temporal features, fuse and generate a spatio-temporal conflict heat map and mark high-risk spatio-temporal coordinates Early warning feedback module: Generate hierarchical early warning signals based on the peak intensity of the spatio-temporal conflict heat map and the distribution density of high-risk spatio-temporal coordinates, and dynamically update the asymmetric time window template parameters in the conflict type template library through early warning accuracy feedback.

Images