A multi-fusion warning method for vehicle-pedestrian collision risk at non-signaled intersections in a V2X environment

Abstract

The present invention discloses a multi-fusion warning method for motor vehicle-pedestrian collision risk at non-signaled intersections in a V2X environment. The present invention focuses on the potential collision warning problem between vehicles and pedestrians in non-signaled intersection scenarios, introduces multi-fusion warning, enhances the prediction efficiency and robustness of pedestrian trajectories, and thus significantly improves the accuracy of detection and warning in non-signaled intersection environments: The present invention integrates the Monte Carlo algorithm, Social GAN, Attention mechanism, and a multi-dimensional warning framework of elliptical buffer domain, significantly improving the detection capability of potential collisions. In response to the uncertainty of pedestrian motion trajectory, the Multi-HeadAttention mechanism is used to optimize the Social GAN ​​model, enhance the time series prediction capability, and solve the problem of partial gradient disappearance, making pedestrian trajectory prediction more accurate. In order to solve the problem of balancing the complexity and accuracy of the warning algorithm, the K value of the buffer area is dynamically adjusted to achieve the duality of the algorithm function, reduce computational redundancy, and ensure the accurate identification of the collision point.

Description

Technical Field

[0001] The present invention relates to the field of traffic accident technology, and in particular to a multi-fusion early warning method for motor vehicle-pedestrian collision risk at an unsignaled intersection in a V2X environment. Background Art

[0002] With the rapid development of modern industry, the number of cars on the road continues to climb worldwide. Frequent traffic accidents have had a profound impact on the economy and society, making research on traffic accident prevention increasingly urgent. This is especially true at intersections without signal controls, where drivers' visibility is obstructed, vehicles travel at high speeds, and the frequent crossing of non-motor vehicles and pedestrians, coupled with their uncertain trajectories, lead to frequent traffic accidents. To address the intense conflicts between motor vehicles, non-motor vehicles, and pedestrians at intersections without signal controls, it is necessary to explore new methods for intersection conflict warning, hoping to significantly reduce the accident rate and further improve traffic safety.

[0003] Intersection Collision Warning (ICW) technology is a key component of active automotive safety technology. Early collision warning systems primarily relied on radar and cameras to perceive the surrounding environment and issue collision warnings to the driver. However, limitations of traditional sensors, such as obstruction by obstacles, have limited the effectiveness and further development of collision warning systems. In recent years, the rise of technologies such as vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) has enabled comprehensive, real-time information sharing among traffic participants, creating a new research hotspot for collision risk identification and warning.

[0004] Collision risk at unsignalized intersections is essentially a matter of motion conflict between traffic participants. Commonly used conflict risk identification metrics are categorized as spatiotemporal conflict indices and probabilistic conflict indices. Early warning methods based on spatiotemporal conflict indices predict potential collision risks by evaluating the temporal and spatial relationships between vehicles. Time to Collision (TTC) is a commonly used evaluation metric. For example, Miller et al. proposed an intersection collision detection method based on TTC, using vehicle TTC for risk assessment. Lan Liuting et al. used TTC, post encroachment time (PET), and deceleration to safety time (DST) as evaluation metrics for conflicts between right-turning vehicles and pedestrians. They found that TTC and PET were negatively correlated with conflict severity, while they were positively correlated with DST. Guo Ruijun et al. proposed an improved TTC calculation method, optimizing the use of rectangular and elliptical models in conflict detection based on geometric models, improving computational accuracy and real-time performance. Furthermore, safety distance warning methods assess risk by analyzing changes in vehicle motion distance. Han Yong et al. proposed a longitudinal and lateral TTC difference model for collision scenarios between cars and electric two-wheelers, optimizing the triggering strategy of the AEB system, improving collision avoidance and reducing collision damage. Li Xingjia et al. proposed a positioning fusion algorithm based on the unscented Kalman filter, which improved positioning accuracy and vehicle safety in the event of positioning failure. Zhao Rui et al. proposed an early warning method based on V2I communication that combines dynamic detection of collision (DDTC) and collision avoidance with circular trajectory (CATC), reducing the rates of missed and false alarms.

[0005] Although early warning methods based on spatiotemporal conflict indicators offer the advantages of low computational complexity and ease of operation, they rely on a single indicator (such as TTC or PET) and are therefore inadequate for complex traffic scenarios. Warning methods based on probabilistic conflict indicators, on the other hand, introduce uncertainty and probabilistic models, assessing the collision risk of vehicles or pedestrians by quantifying them. For example, Laugier et al. used hidden Markov models (HMMs) and Gaussian processes to identify vehicle behavior and predict future collision risk, validating their effectiveness in uncertain environments. Joerer et al. further improved detection accuracy by assessing the collision risk of all possible vehicle trajectories using acceleration probability distributions. Berthelot et al. proposed a method for calculating the TTC probability distribution based on uncertain inputs, enhancing system reliability by addressing input noise. Regarding pedestrian collision risk, Peng Liqun et al. proposed a random geometric model of the pedestrian collision zone based on vehicle-to-vehicle (V2P) communication technology. Taking into account factors such as communication delay and positioning error, they established a pedestrian-vehicle collision probability model and verified its accuracy in complex scenarios through simulation experiments. Yang Xu et al. constructed a real-time early warning model for pedestrian-vehicle collision risk based on catastrophe theory, combining Binary Logistic regression with the catastrophe progression method for evaluation. Jin Yuanyuan et al. proposed an early warning method based on the probability of vehicle collision at intersections. This method calculates the collision probability by predicting future vehicle behavior and validates its effectiveness in simulations.

[0006] In summary, previous research on intersection collision warning has largely failed to fully consider the complexity and variability of pedestrian behavior, overlooking the challenges posed by pedestrian erratic movements. Therefore, it is necessary to improve existing technologies and provide solutions focused on the potential collision warning problem between vehicles and pedestrians at unsignalized intersections. Summary of the Invention

[0007] Therefore, based on the above background, the present invention provides a multi-fusion warning method for motor vehicle-pedestrian collision risk at non-signaled intersections in a V2X environment, focusing on the potential collision warning problem between vehicles and pedestrians in non-signaled intersection scenarios. By introducing multi-fusion warning technology, the prediction efficiency and robustness of pedestrian trajectories are enhanced, thereby significantly improving the accuracy of detection and warning in non-signaled intersection environments, providing strong support for comprehensively improving road safety.

[0008] The technical solution provided by the present invention is:

[0009] A multi-fusion warning method for vehicle-pedestrian collision risk at an unsignaled intersection in a V2X environment comprises the following steps:

[0010] 1) Collecting and sharing information about motor vehicles and pedestrians; the information includes movement parameters and real-time locations of motor vehicles and pedestrians;

[0011] 2) Input the information of motor vehicles and pedestrians into the vehicle trajectory model and pedestrian trajectory model respectively to calculate the future movement trajectory of the vehicle and pedestrian;

[0012] 3) Based on the future motion trajectories of vehicles and pedestrians, an elliptical buffer algorithm is used to preliminarily screen pedestrians and identify potential collision targets within the buffer zone.

[0013] 4) The historical trajectory of the identified potential collision target is input into the noisy pedestrian trajectory model to generate a large number of similar predicted trajectories;

[0014] 5) Based on the predicted trajectory generated in step 4), using the elliptical buffer algorithm, adjusting the parameter K of the elliptical buffer, calculating and storing the potential collision point, potential collision time, and potential collision distance between the vehicle and the potential collision target;

[0015] 6) Based on the potential collision point, potential collision time, and potential collision distance obtained in step 5), a Monte Carlo algorithm is used to calculate the collision expectation value and collision probability. Combined with the time (TTM) from the driver's awareness of the danger to the vehicle's complete stop, a joint collision probability is calculated based on a joint probability formula;

[0016] 7) Compare the combined collision probability obtained in step 6) with a preset warning threshold. If the combined collision probability exceeds the preset threshold, the collision warning system is immediately triggered to alert the driver.

[0017] Furthermore, the motion parameters of the motor vehicle in step S1 include travel speed, travel direction (heading angle), and angular velocity (steering angular velocity); the motion parameters of the pedestrian include travel speed, travel direction (heading angle), and steering angle;

[0018] The real-time positions of the motor vehicles and pedestrians include GPS coordinates or roadside relative coordinates.

[0019] Furthermore, the vehicle trajectory model in step S2 is a state vector P that can update the vehicle's driving direction and position information in real time based on a two-dimensional plane kinematic model. V (t), used to predict the future trajectory of the vehicle in the plane;

[0020] The state vector P V (t) See formula (1):

[0021] P V (t)=[x(t),y(t),v x (t),v y (t)] T (1)

[0022] In formula (1), x(t) and y(t) represent the position coordinates of the vehicle on the two-dimensional plane at time t. The formulas for x(t) and y(t) are shown in the following formula (2):

[0023]

[0024] x(t-1) and y(t-1) represent the position of the vehicle at the previous moment t-1, and the velocity component of the vehicle at moment t is represented by v x (t) and v y (t) indicates that P V Corresponding to the vehicle's speed in the x and y directions respectively.

[0025] Furthermore, the pedestrian trajectory model in step S2 is constructed based on the Social GAN ​​model, which incorporates a multi-head attention mechanism into the generator input and introduces a diversity loss function.

[0026] Furthermore, the objective function of the adversarial neural network of the pedestrian trajectory model is shown in the following formula (3):

[0027]

[0028] Its core is to generate realistic future trajectories through adversarial training between the generator G and the discriminator D. The generator G tries to generate a trajectory G(z) that can "fool" the discriminator D, while the discriminator D tries to distinguish between the real trajectory x and the generated trajectory G(z);

[0029] In the multi-head attention mechanism, the relationship between each time step is calculated as shown in equations (4) to (6):

[0030]

[0031] Z=MultiHead(Q,K,V)(5)

[0032] G(z)=Generator(z,Z)(6)

[0033] Where Q, K, and V are query, key, and value matrices respectively, T is the transpose, and d k is the dimension of the key; Z is the output of the multi-head self-attention mechanism, G(z) is the trajectory generator, and z represents the latent variable.

[0034] Autonomous attention loss L introduced in pedestrian trajectory model att and diversity loss L var , see the following equations (7) and (8):

[0035]

[0036]

[0037] In the above formula (7), A gen Represents the self-attention matrix generated by the model, A real represents the true self-attention matrix of the model; in formula (8), Y pred (k) Represents the k-th predicted trajectory. pred (k ' ) Represents the k'th predicted trajectory, which is another trajectory generated by the model and is the same as Y pred (k) There are differences between them.

[0038] The total loss function of the adversarial neural network is shown in the following formula (9):

[0039] L total =L GAN +λ1L att +λ2L var (9)

[0040] Where λ1 and λ2 are hyperparameters, L GAN is the adversarial loss function.

[0041] Furthermore, in step S4, the prediction trajectory formula generated after introducing noise into the pedestrian trajectory model is shown in the following formula (10):

[0042]

[0043] Where, represents the predicted trajectory of the i-th pedestrian of the k-th potential collision target, H k represents the historical trajectory of the input, z is a random noise vector, N(0,σ 2 ) is a Gaussian noise term used to perturb the trajectory at each generation to increase diversity.

[0044] Furthermore, the basic formula of the buffer domain algorithm in steps 3) and 5) is shown in the following formula (11):

[0045]

[0046] Where x c and y c are the center points of the ellipse, a and b are the major and minor axes of the ellipse, θ is the rotation angle of the ellipse; x and y are the coordinates of the pedestrian or vehicle target point, respectively;

[0047] The major axis and minor axis of the algorithm are defined as follows:

[0048]

[0049]

[0050] d b =K·v vul ·t rea (14)

[0051] Where L and W are the length and width of the vehicle respectively, d b is the buffer distance used to extend the physical boundary of the vehicle; v vul is the speed of the pedestrian, t rea is the driver's reaction time, and K is the adjustment coefficient.

[0052] Furthermore, in step 3), pedestrians are preliminarily screened using the buffer domain algorithm using the following formula (15):

[0053]

[0054] Specifically, the four vertices of the pedestrian rectangle are substituted into the rotation ellipse equation of formula (15) to determine whether it is inside the ellipse;

[0055] In step 5), the potential collision point, potential collision time, and potential collision distance between the vehicle and the potential collision target are calculated using the following equations (16) and (17):

[0056] y j =m j x+e j ,j=1,2,3,4(16)

[0057]

[0058] In the above formula, j represents the four sides of the rectangle, where the upper and lower, left and right boundaries are each a straight line; m j is the slope of the jth edge, e j is the intercept of the jth side;

[0059] Specifically, the four vertices and edges of the pedestrian rectangle are substituted into the geometric equation of the rotation ellipse in equation (16) (17).

[0060] Furthermore, the joint probability formula in step 6) is shown in the following formula (24):

[0061] P(t≤TTM,TTC∈R)=P(t≤TTM|TTC∈R)P(TTC∈R)(24)

[0062] The TTC in the above formula represents the time required for a traffic participant to collide. The joint probability P(t≤TTM,TTC∈R) reflects the situation in which a collision occurs (TTC∈R) and the collision time t is less than TTM in all samples.

[0063] The conditional probability P(t≤TTM / TTC∈R) is used to describe the possibility that the collision time t is less than TTM under the condition that the collision has already occurred (that is, TTC∈R); the collision probability P(TTC∈R) represents the proportion of TTC in all samples that falls into the risk interval R;

[0064] The expected value E(d) of the collision distance distribution and the collision probability P(TTC∈R) calculated by Monte Carlo are shown in the following equations (20) to (22):

[0065] N total =m*h(20)

[0066]

[0067]

[0068] In the above formula (20), h represents the number of potential targets, m represents the number of future trajectories generated by the same potential target; in formula (21), d k represents the collision distance of the kth potential target; N in formula (22) total is the total number of sampling times of the Monte Carlo algorithm, and II is the indicator function used to indicate whether there is a collision;

[0069] In formula (22) is a binary variable representing the collision judgment result, as shown in formulas (18) and (19):

[0070]

[0071]

[0072] Where, represents the i-th trajectory of the k-th potential target, m and h are natural integers, and N represents the number of time steps in the trajectory; The elliptical buffer domain formula (16) is used to determine whether the kth potential collision target collides with the vehicle trajectory in the i-th predicted trajectory;

[0073] The calculation of TTM is shown in the following formula (23):

[0074] TTM=t rea +t brak (twenty three)

[0075] Where h represents the number of potential targets, m represents the number of future trajectories generated by the same potential target, and N total is the total number of Monte Carlo sampling, TTM is the driver and vehicle reaction time t rea and braking time t brak The sum of .

[0076] Based on the same inventive concept, the present invention also provides an electronic device, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the steps of the above method when executing the program.

[0077] The above technical solution has the following beneficial effects:

[0078] This paper focuses on the potential collision warning problem between vehicles and pedestrians in unsignalized intersections. It introduces multi-fusion warning to enhance the prediction efficiency and robustness of pedestrian trajectories, thereby significantly improving the accuracy of detection and warning in unsignalized intersection environments:

[0079] To address the complexity of motor vehicle-pedestrian collision scenarios, the present invention integrates the Monte Carlo (MC) algorithm, Social GAN, the Attention mechanism, and a multi-dimensional warning framework of elliptical buffer domains, which can significantly improve the detection capability of potential collisions.

[0080] In order to address the uncertainty of pedestrian motion trajectories, the Multi-HeadAttention mechanism is used to optimize the SocialGAN model, which not only enhances the time series prediction capability, but also effectively solves the problem of partial gradient disappearance, making pedestrian trajectory prediction more accurate.

[0081] In order to solve the balance problem between the complexity and accuracy of the early warning algorithm, the present invention introduces an elliptical buffer domain algorithm, which uses the K value of the dynamic adjustment buffer area to achieve the duality of the algorithm function, reduce calculation redundancy, and ensure the accurate identification of the collision point.

[0082] Simulations have shown that in the scenario of unsignalized intersections, the accuracy of the present invention in warning of motor vehicle-pedestrian collisions is as high as 94.5%, which is significantly better than the traditional TTC model and braking distance model. It also greatly reduces the missed alarm rate and false alarm rate, and can provide a reliable basis for real-time vehicle warnings and the implementation of avoidance measures, providing strong protection for pedestrian traffic safety at unsignalized intersections. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Attachment Figure 1 This is an example of a non-signaled intersection scenario in an embodiment.

[0084] Attachment Figure 2 This is an example of an unsignaled intersection target recognition scenario in an embodiment.

[0085] Attachment Figure 3 YOLOv5s network architecture for an embodiment.

[0086] Attachment Figure 4FIG. 4 is a vehicle motion trajectory diagram of an embodiment.

[0087] Attachment Figure 5 This is a graph showing the change in the loss function of the training set of the pedestrian trajectory model of the present invention.

[0088] Attachment Figure 6 This is the network framework for pedestrian trajectory prediction of the present invention.

[0089] Attachment Figure 7 This is a flow chart of the multi-fusion early warning method of the present invention.

[0090] Attachment Figure 8 Schematic diagram of collision detection using the elliptical buffer domain algorithm of the present invention.

[0091] Attachment Figure 9 The Monte Carlo algorithm of the present invention utilizes a pedestrian trajectory model to predict trajectories of potential targets.

[0092] Attachment Figure 10 Graph showing the collision probability changes under different thresholds in the embodiment.

[0093] Attachment Figure 11 is the collision time distribution of the embodiment.

[0094] Attachment Figure 12 is the collision distance distribution of the embodiment.

[0095] Attachment Figure 13 1s simulation accident analysis diagram of the embodiment, wherein (a) T=1s real collision scene; T=1s simulation collision scene; (c) T=1s joint collision probability change diagram.

[0096] Attachment Figure 14 These are simulation accident analysis diagrams for T=1.45s in the embodiment, where (a) T=1.45s is a real collision scene; (b) T=1.45s is a simulation collision scene; and (c) T=1.45s is a combined collision probability change diagram.

[0097] Attachment Figure 15 These are the simulation accident analysis diagrams for T=4.2s of the embodiment, where (a) the real collision scene at T=4.2s; (b) the collision simulation scene at T=4.2s without warning; (c) the joint collision probability change diagram at T=4.2s warning; and (d) the collision simulation scene at T=4.2s warning. DETAILED DESCRIPTION

[0098] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0099] The technical solution of the present invention:

[0100] A multi-fusion warning method for vehicle-pedestrian collision risk at an unsignaled intersection in a V2X environment comprises the following steps:

[0101] 1) Using V2X communication technology, roadside units (RSUs) collect and share information about vehicles and pedestrians; the information includes their motion parameters and real-time locations.

[0102] Through V2X communication technology, vehicles exchange data with RSUs via OBUs (on-board units). RSUs aggregate and share information with all traffic participants in the area; the information includes the movement parameters and real-time locations of vehicles and pedestrians.

[0103] The motion parameters of the motor vehicle in step S1 include driving speed, driving direction (heading angle), and angular velocity (steering angular velocity); the motion parameters of the pedestrian include driving speed, driving direction (heading angle), and steering angle;

[0104] The real-time positions of the motor vehicles and pedestrians include GPS coordinates or roadside relative coordinates.

[0105] 2) Using the vehicle trajectory model and the pedestrian trajectory model to obtain the future motion trajectories of the vehicle and pedestrian based on the information in step 1);

[0106] The vehicle trajectory model in this step is a state vector P that can update the vehicle's driving direction and position information in real time based on a two-dimensional plane kinematic model. V (t), used to predict the future trajectory of the vehicle in the plane;

[0107] The state vector P V (t) See formula (1):

[0108] P V (t)=[x(t),y(t),v x (t),v y (t)] T (1)

[0109] In formula (1), x(t) and y(t) represent the position coordinates of the vehicle on the two-dimensional plane at time t. The formulas for x(t) and y(t) are shown in the following formula (2):

[0110]

[0111] x(t-1) and y(t-1) represent the position of the vehicle at the previous moment t-1, and the velocity component of the vehicle at moment t is represented by v x (t) and v y (t) indicates that P V They correspond to the speed of the vehicle in the x and y directions respectively, and △t represents the time interval.

[0112] The pedestrian trajectory model is constructed based on the Social GAN ​​model, which incorporates a multi-head attention mechanism into the generator input and introduces a diversity loss function.

[0113] The objective function of the adversarial neural network of the pedestrian trajectory model is shown in the following formula (3):

[0114]

[0115] Its core is to generate realistic future trajectories through adversarial training between the generator G and the discriminator D. The generator G tries to generate a trajectory G(z) that can "fool" the discriminator D, while the discriminator D tries to distinguish between the real trajectory x and the generated trajectory G(z);

[0116] In the multi-head attention mechanism, the relationship between each time step is calculated as shown in equations (4) to (6):

[0117]

[0118] Z=MultiHead(Q,K,V)(5)

[0119] G(z)=Generator(z,Z)(6)

[0120] Where Q, K, and V are query, key, and value matrices respectively, T is the transpose, and d k is the dimension of the key; Z is the output of the multi-head self-attention mechanism, G(z) is the trajectory generator, and z represents the latent variable.

[0121] Autonomous attention loss L introduced in pedestrian trajectory model att and diversity loss L var , see the following equations (7) and (8):

[0122]

[0123]

[0124] In the above formula (7), A gen Represents the self-attention matrix generated by the model, A real represents the true self-attention matrix of the model; in formula (8), Y pred (k) Represents the k-th predicted trajectory. pred (k ' ) Represents the k'th predicted trajectory, which is another trajectory generated by the model and is the same as Y pred (k) There are differences between them.

[0125] The total loss function of the adversarial neural network is shown in the following formula (9):

[0126] L total =L GAN +λ1L att +λ2L var (9)

[0127] Where λ1 and λ2 are hyperparameters, L GAN is the adversarial loss function.

[0128] 3) Based on the future motion trajectories of the vehicle and pedestrians obtained in step 2), the elliptical buffer algorithm is used to preliminarily screen the pedestrians and identify potential collision targets within the buffer;

[0129] The basic formula of the buffer domain algorithm (rotated ellipse equation) is shown in the following formula (11):

[0130]

[0131] Where x c and y c are the center points of the ellipse, a and b are the major and minor axes of the ellipse, θ is the rotation angle of the ellipse; x and y are the coordinates of the pedestrian or vehicle target point, respectively;

[0132] The major axis and minor axis of the algorithm are defined as follows:

[0133]

[0134]

[0135] d b =K·v vul ·t rea (14)

[0136] Where L and W are the length and width of the vehicle respectively, d b is the buffer distance used to extend the physical boundary of the vehicle; vvul is the speed of the pedestrian, t rea is the driver's reaction time, and K is the adjustment coefficient.

[0137]

[0138] Specifically, the four vertices of the pedestrian rectangle are substituted into the rotation ellipse equation of formula (15) to determine whether it is inside the ellipse.

[0139] 4) Introducing noise into the pedestrian trajectory model and inputting the historical trajectories of the potential collision targets identified in step 3) to generate a large number of similar predicted trajectories;

[0140] In this step, the prediction trajectory formula generated by introducing noise into the pedestrian trajectory model is shown in the following formula (10):

[0141]

[0142] Where, represents the predicted trajectory of the i-th pedestrian of the k-th potential collision target, H k represents the historical trajectory of the input, z is a random noise vector, N(0,σ 2 ) is a Gaussian noise term used to perturb the trajectory at each generation to increase diversity.

[0143] 5) Based on the predicted trajectory generated in step 4), using the elliptical buffer algorithm, adjusting the parameter K of the elliptical buffer, calculating and storing the potential collision point, potential collision time, and potential collision distance between the vehicle and the potential collision target;

[0144] In this step, the potential collision point, potential collision time, and potential collision distance between the vehicle and the potential collision target are calculated using the following equations (16) and (17):

[0145] y j =m j x+e j ,j=1,2,3,4(16)

[0146]

[0147] In the above formula, j represents the four sides of the rectangle, where the upper and lower, left and right boundaries are each a straight line; m j is the slope of the jth edge, e j is the intercept of the jth side;

[0148] Specifically, the four vertices and edges of the pedestrian rectangle are substituted into the geometric equation of the rotation ellipse in equation (16) (17).

[0149] 6) Based on the potential collision point, potential collision time, and potential collision distance obtained in step 5), a Monte Carlo algorithm is used to calculate the collision expectation value and collision probability. Combined with the time (TTM) from the driver's awareness of the danger to the vehicle's complete stop, a joint collision probability is calculated based on a joint probability formula;

[0150] The joint probability formula in this step is shown in the following formula (24):

[0151] P(t≤TTM,TTC∈R)=P(t≤TTM|TTC∈R)P(TTC∈R)(24)

[0152] The TTC in the above formula represents the time required for a traffic participant to collide. The joint probability P(t≤TTM,TTC∈R) reflects the situation in which a collision occurs (TTC∈R) and the collision time t is less than TTM in all samples.

[0153] The conditional probability P(t≤TTM / TTC∈R) is used to describe the possibility that the collision time t is less than TTM under the condition that the collision has already occurred (that is, TTC∈R); the collision probability P(TTC∈R) represents the proportion of TTC in all samples that falls into the risk interval R;

[0154] The expected value E(d) of the collision distance distribution and the collision probability P(TTC∈R) calculated by Monte Carlo are shown in the following equations (20) to (22):

[0155] N total =m*h(20)

[0156]

[0157]

[0158] In the above formula (20), h represents the number of potential targets, m represents the number of future trajectories generated by the same potential target; in formula (21), d k represents the collision distance of the kth potential target; N in formula (22) total is the total number of sampling times of the Monte Carlo algorithm, and II is the indicator function used to indicate whether there is a collision;

[0159] In formula (22), C i k is a binary variable representing the collision judgment result, as shown in formulas (18) and (19):

[0160]

[0161]

[0162] Where, represents the i-th trajectory of the k-th potential target, m and h are natural integers, and N represents the number of time steps in the trajectory; The elliptical buffer domain formula (16) is used to determine whether the kth potential collision target collides with the vehicle trajectory in the i-th predicted trajectory;

[0163] The calculation of TTM is shown in the following formula (23):

[0164] TTM=t rea +t brak (twenty three)

[0165] Where h represents the number of potential targets, m represents the number of future trajectories generated by the same potential target, and N total is the total number of Monte Carlo sampling, TTM is the driver and vehicle reaction time t rea and braking time t brak The sum of .

[0166] 7) Compare the combined collision probability obtained in step 6) with a preset warning threshold. If the combined collision probability exceeds the preset threshold, the collision warning system is immediately triggered to alert the driver.

[0167] Example 1: This example further illustrates the present invention.

[0168] (1) Construction of vehicle trajectory model and pedestrian trajectory model

[0169] First, the trajectory models of vehicles and pedestrians are constructed.

[0170] In unsignalized intersections, this paper uses a kinematic model to construct a vehicle trajectory model, given the strictness of road traffic regulations and the relative predictability of driver behavior. In contrast, pedestrian movement, speed, and behavior are subject to greater uncertainty and variability. Therefore, to more accurately capture and predict pedestrian movement paths, this paper employs an adversarial neural network to construct a pedestrian trajectory model.

[0171] At the same time, in order to build a widely applicable pedestrian trajectory prediction model for unsignaled intersections, this paper uses a drone (DJI Mavic 3) to collect field data at multiple unsignaled intersections in Guilin (see Figure 1 By applying the YOLOv5x algorithm, this embodiment extracts valid trajectory information of 350 pedestrians from the collected video data (see Figure 2 Unsignaled intersection target recognition scenario example), target trajectory information after feature extraction (see Figure 3The dataset used for training and testing in this study is a YOLOv5s network architecture. During data collection, images of pedestrians on the road are captured at 10-frame intervals, including the time frame number, pedestrian ID, and X and Y coordinate information.

[0172] ①Vehicle trajectory model

[0173] The present invention adopts a simplified two-dimensional plane kinematic model (see Figure 4 ), which is used to describe the motion state of a vehicle at an unsignaled intersection, including characteristics such as displacement, velocity, and acceleration. Based on data sharing of V2X communication technology, the vehicle's motion parameters (including position, velocity, heading angle, etc.) are extracted in real time through RSU, providing key input for collision warning. Using the vehicle motion parameters, the state vector P is constructed based on the kinematic model. V (t) is used to update the vehicle’s direction and position information in real time, thereby predicting the vehicle’s future trajectory within the plane. See Formula 1-2 for details.

[0174] P V (t)=[x(t),y(t),v x (t),v y (t)] T (1)

[0175]

[0176] Where x(t) and y(t) represent the position coordinates of the vehicle at time t on the two-dimensional plane, and x(t-1) and y(t-1) represent the position of the vehicle at the previous time t-1. The velocity component of the vehicle at time t is given by v x (t) and v y (t) indicates that P V Corresponding to the vehicle's speed in the x and y directions respectively.

[0177] ② Pedestrian trajectory model

[0178] This paper, based on the Social GAN ​​model, effectively simulates the social interactions between pedestrians through a social pooling layer and uses pedestrian historical trajectory data to predict future trajectories. However, traditional Social GAN ​​models have limitations in capturing complex temporal dependencies and suffer from the vanishing gradient problem.

[0179] Therefore, the present invention incorporates a multi-head attention mechanism into the decoder of the Social GAN ​​model. The multi-head attention mechanism has the function of enhancing the temporal modeling capability of the original model and solving the gradient vanishing problem. It can simultaneously focus on multiple key time steps in the historical trajectory data and dynamically assign attention weights, thereby accurately extracting information that is crucial for future trajectory prediction.

[0180] The optimized model uses the previous eight time steps of historical trajectory data as input to the generator, combined with a multi-head attention mechanism to capture the pedestrian's motion trends and potential intentions. By processing the output of the attention layer through an LSTM encoder, the model learns the long-term dependencies of the pedestrian's historical trajectory and generates trajectory predictions for the next 12 time steps based on this.

[0181] Building on this foundation, the optimized model indirectly combines a social pooling layer and a multi-head attention mechanism via an LSTM encoder. The social pooling layer models the complex interactions between pedestrians, while the multi-head attention mechanism focuses on capturing temporal dependencies in historical trajectories. The LSTM encoder integrates the output features from both modules to generate a comprehensive latent state representation. This not only effectively captures the interactions between pedestrians but also significantly improves the model's accuracy in handling temporal dependencies, thereby achieving better performance in trajectory prediction.

[0182] Formula (3) describes the objective function of its adversarial network, the core of which is to generate realistic future trajectories through adversarial training between the generator G and the discriminator D. The generator G tries to generate a trajectory G(z) that can "fool" the discriminator D, while the discriminator D tries to distinguish between the real trajectory x and the generated trajectory G(z).

[0183]

[0184] In the multi-head attention mechanism, the mutual relationship of each time step is calculated, as shown in Formula 4-6.

[0185]

[0186] Z=MultiHead(Q,K,V)(5)

[0187] G(z)=Generator(z,Z)(6)

[0188] Where Q, K, and V are query, key, and value matrices, respectively, and d k The dimension of the key.

[0189] Z is the context information obtained through the multi-head attention mechanism, z represents the latent variable, and the generator G(z) uses this information to generate a prediction trajectory that is more consistent with the temporal correlation. In order to deal with the multimodal problem in trajectory prediction, formula 7-8 introduces the autonomous attention loss L att And the diversity loss function L var , which is calculated as follows:

[0190]

[0191]

[0192] In the above formula (7), A gen Represents the self-attention matrix generated by the model, A real represents the true self-attention matrix of the model; in formula (8), Y pred (k) Represents the k-th predicted trajectory. pred (k ' ) Represents the k'th predicted trajectory, which is another trajectory generated by the model and is the same as Y pred (k) There are differences between them.

[0193] This loss function encourages the generation of more diverse trajectories by selecting the one that is farthest from the true trajectory among the multiple candidate trajectories generated. Finally, Formula 9 shows the total loss function L of the adversarial neural network. total , and the training set L total Change curve see Figure 5 .

[0194] L total =L GAN +λ1L att +λ2L var (9)

[0195] Where λ1 and λ2 are hyperparameters that control the importance of each loss term. GAN Supervise the authenticity of the generated trajectory, the self-attention mechanism loss function L att The design of should consider the temporal correlation and spatial dependency between the generated trajectory and the real trajectory, while the diversity loss function L var By encouraging the generation of diverse trajectories, we avoid generating overly similar trajectories. Furthermore, to enable the model to better cope with interactive scenarios at unsignaled intersections, we introduce a balance between diversity loss and multimodal trajectories, thereby improving both the accuracy and diversity of trajectory prediction.

[0196] Figure 6This is a network framework for pedestrian trajectory prediction, combining the architectures of Multi-HeadAttention and Social GAN. It is used to generate multiple predicted trajectories for vulnerable traffic groups, such as single pedestrians, from a speed-trajectory time series dataset. Model Optimization: The model input data includes the target's historical trajectory, speed, and other dynamic features as the model's foundational information. First, the Multi-HeadAttention module captures the interactive relationships between the target's basic information and generates a context vector. Subsequently, the attention mechanism outputs the vector through the Add & Norm operation, resulting in a normalized context vector to ensure information stability and model numerical stability. The context vector obtained through the Add & Norm operation is then input into the LSTM encoder to extract the target's historical motion features. The Pooling module aggregates the interactive information of multiple targets to generate a global feature representation. Based on this, the LSTM decoder recursively generates trajectories for multiple future time steps based on these features, outputting multiple diverse predicted trajectories for each target.

[0197] Furthermore, to further explore the diversity of pedestrian trajectory prediction, noise is introduced into the generation process of the pedestrian trajectory prediction model, enabling the generation of multiple similar, but not identical, predicted trajectories based on a single historical trajectory. The injection of noise increases the randomness and uncertainty of trajectory generation, enabling the model to not only capture trends and patterns in historical trajectories but also explore the various possibilities of trajectory evolution. Specifically, by adding noise to the output of the generator, the model can generate multiple similar, but not identical, predicted trajectories based on the same historical trajectory, as shown in Equation 10. Through this mechanism, the optimization model can provide the MC algorithm (Monte Carlo algorithm) with multiple random samples of pedestrian trajectories, increasing the algorithm's flexibility and reliability.

[0198]

[0199] Where, represents the Monte Carlo random sampling of the kth potential collision target i-th pedestrian predicted trajectory, H k Represents the historical trajectory of the input, z usually represents a random noise vector. 2 ) is a Gaussian noise term used to perturb the trajectory at each generation to increase diversity.

[0200] (2) Collision detection and warning

[0201] Based on the vehicle and pedestrian trajectory models constructed above, the following steps are mainly included to implement multi-fusion warning in the V2X communication environment: screening of potential collision targets within the elliptical buffer area, joint collision probability assessment combining MC and TTM (Total Time to Maneuver), and collision warning threshold determination.

[0202] The multi-fusion warning method first uses information sharing between roadside units (RSUs) and combines vehicle and pedestrian trajectory prediction models to determine the future motion trajectories of vehicles and pedestrians. An elliptical buffer algorithm is then used to preliminarily screen pedestrians and identify potential collision targets within the buffer. After identifying potential collision targets, the pedestrian trajectory prediction model generates a large number of similar predicted trajectories for these targets, rather than using a traditional random sampling mechanism. The elliptical buffer parameter K is then adjusted and geometric equations (see Formula 17) are used to calculate and store the future collision point and distance (potential collision point, potential collision time, and potential collision distance) between the vehicle and the potential collision target. A Monte Carlo algorithm is then used to calculate the expected values ​​of the collision probability, collision time, and collision distance for the collision data. The joint collision probability is then calculated based on the joint probability formula (see Formula 24), taking into account the time from driver awareness of the danger to complete vehicle stop (TTM). Finally, the calculated joint collision probability is compared with a preset warning threshold. If the joint collision probability exceeds the preset threshold, the collision warning system is immediately triggered, alerting the driver so that they can take necessary evasive action. The process of the entire multi-fusion early warning method is shown in Figure 7 .

[0203] ① Ellipse buffer domain algorithm

[0204] This paper combines the concept of buffer zone with the geometric collision model, introduces an adjustable parameter K, and dynamically adjusts the major and minor axes of the elliptical buffer zone based on pedestrian speed, driver reaction time, and vehicle information. It constructs an elliptical buffer zone algorithm, aiming to more efficiently detect and screen potential collision risks.

[0205] The elliptical buffer algorithm expands the collision detection boundary of the vehicle by introducing a buffer area and screening out potential collision targets. At the same time, the algorithm realizes dual functions based on the adjustment of K value, one is used for preliminary screening and identification of potential collision targets, and the other is used for accurate determination of collision points. Figure 8 As shown, the yellow elliptical area represents the target motor vehicle, and the blue rectangle represents pedestrians or other vulnerable traffic groups.

[0206] The basic formula of the elliptical buffer domain algorithm is 11:

[0207]

[0208] Where x c and y c are the center points of the ellipse, a and b are the major and minor axes of the ellipse, and θ is the rotation angle of the ellipse.

[0209] The major axis and minor axis of the algorithm are defined as follows (12)-(14):

[0210]

[0211]

[0212] d b =K·v vul ·t rea (14)

[0213] Where L and W are the length and width of the vehicle respectively, d b is the buffer distance used to extend the physical boundary of the vehicle. vul is the speed of the pedestrian, t rea is the driver's reaction time, and K is the adjustment coefficient.

[0214] The elliptical buffer algorithm is used to filter potential collision targets:

[0215] Using the above formula, the elliptical buffer algorithm sets a relatively large K value during the initial screening phase to expand the elliptical buffer. This allows for wider coverage of surrounding traffic participants, ensuring that all potential threats are identified. Specifically, the algorithm first uses vehicle and pedestrian information collected by roadside units (RSUs) to predict their future trajectories and compares them with the expanded elliptical buffer to screen out potential collision targets.

[0216] In the initial screening phase, this paper uses the geometric relationship between ellipses and rectangles to quickly identify potential collision targets using inequality conditions. Specifically, the four vertices of the pedestrian rectangle are substituted into the equation of the rotated ellipse to determine whether it lies inside the ellipse (see Equation 15).

[0217]

[0218] In the formula, (x c ,y c ) is the coordinate of the center of the ellipse, a and b are the major and minor axes respectively, θ is the rotation angle of the ellipse, (x i ,y i ) are the coordinates of the vertices of the rectangle.

[0219] In the subsequent joint collision probability calculation stage of the elliptical buffer domain algorithm, the K value is reduced to narrow the range of the elliptical buffer domain, and the geometric equation is used to accurately calculate the collision time and collision distance at different time points, providing reliable data support for the subsequent joint collision probability assessment.

[0220] During the joint collision probability calculation phase, this paper reduces the K value to narrow the elliptical buffer zone, utilizing it for more accurate collision calculations based on the Monte Carlo algorithm. By substituting the rectangle's vertices and edges (see Equation 16) into the geometric equation of the rotated ellipse (Equation 17), the potential collision point, collision time, and collision distance between the vehicle and the target traffic participant can be calculated (see Equations 13-14).

[0221] y j =m j x+e j ,j=1,2,3,4(16)

[0222]

[0223] Where y j represents the equation of a straight line, j represents the jth side of a rectangle (the upper and lower, left and right boundaries are each a straight line), m j is the slope of the jth edge, e j is the intercept of the jth side.

[0224] Due to the diversity of traffic participants and the random nature of their behavior, traditional collision detection methods struggle to effectively handle complex interactive scenarios. Therefore, the elliptical buffer algorithm constructed in this paper can improve the warning capability and robustness of traffic at unsignalized intersections by expanding the detection boundary and flexibly adjusting the buffer area. This algorithm is particularly effective in handling group pedestrian behavior, improving the adaptability and effectiveness of the warning algorithm.

[0225] ②Collision probability and warning

[0226] After the potential collision targets are screened out by the elliptical buffer domain algorithm, the scenario where there is a collision risk between the vehicle and the pedestrian is identified. In order to accurately quantify the collision time, distance and probability, the Monte Carlo algorithm is then used as the main analysis tool to combine the trajectories of the vehicle and pedestrian trajectory models, and simulate different trajectory scenarios through multiple random sampling to estimate the collision risk between the vehicle and the pedestrian. Specifically, a unique predicted trajectory is obtained from the vehicle trajectory model to represent the future motion path of the vehicle. Multiple possible trajectories are sampled from the pedestrian trajectory prediction model to simulate the uncertain future motion of pedestrians. In each Monte Carlo simulation, a vehicle trajectory and multiple pedestrian trajectories are randomly selected, and the occurrence of collisions is calculated through collision detection, and the frequency of collisions is counted to finally estimate the collision probability (see Formula 22). In addition, in order to improve the accuracy of collision judgment, the present invention introduces the elliptical buffer domain algorithm as a collision judgment criterion, and combines TTM and joint probability analysis to comprehensively evaluate the collision risk of the vehicle.

[0227] Figure 9The present invention shows that the historical trajectory information of potential collision targets is input into the pedestrian trajectory model, and i pedestrian future trajectory samples are generated for the kth potential target and recorded in (Refer to Formula 18).

[0228] First, in Monte Carlo random sampling In each time step of each predicted trajectory, the elliptical buffer domain algorithm (refer to Formula 17) is used to detect the relative position relationship between the pedestrian and the vehicle, so as to make a collision judgment. The potential conflict point, collision time, and collision distance between the vehicle and pedestrian are recorded and obtained, and the expected value E(d) of the collision distance distribution and the collision probability P(TTC∈R) are calculated using the Monte Carlo algorithm. P(TTC∈R) represents the proportion of TTCs in all samples that fall within the risk interval R, as shown in the following formula.

[0229]

[0230]

[0231] N total =m*h(20)

[0232]

[0233]

[0234] Where m and h are natural integers, and N represents the number of time steps in the trajectory. is the collision judgment function of the kth potential collision target in the i-th predicted trajectory. k represents the collision distance of the kth potential target, and II represents the indicator function, which is used to indicate whether there is a collision. total is the total number of Monte Carlo samplings.

[0235] Then, after considering the influence of TTM (the sum of the driver's reaction time and braking time), the joint collision probability is selected as the final criterion for evaluating the collision risk. The joint collision probability is calculated based on the total result of the above trajectory judgment.

[0236] The conditional probability P(t≤TTM / TTC∈R) is used to describe the possibility that the collision time t is less than TTM under the condition that the collision has already occurred (that is, TTC∈R);

[0237] The joint probability P(t≤TTM, TTC∈R) reflects the situation in which a collision occurs (TTC∈R) and the collision time t is less than TTM in all samples. See formula 23-24 for details:

[0238] TTM=trea +t brak (twenty three)

[0239] P(t≤TTM,TTC∈R)=P(t≤TTM|TTC∈R)P(TTC∈R)(24)

[0240] Where TTM is the reaction time between the driver and the vehicle rea and braking time t brak The sum of .

[0241] ②Collision warning

[0242] In this embodiment, the combined collision probability thresholds of 0.25, 0.55 and 0.95 are selected to conduct a single-vehicle accident simulation experiment. Figure 10 As shown in the figure, when the collision time between the vehicle and the pedestrian is 4.2 seconds, the warning system with a threshold of 0.55 performs best. Therefore, this paper determines the optimal combined collision warning threshold to be 0.55. When the collision probability exceeds this value, the system immediately issues a warning, prompting the driver to take evasive or braking measures to prevent a collision.

[0243] Example 2: This example conducts simulation verification on the present invention.

[0244] To provide early warnings for vehicle collision risks at unsignalized intersections and road sections, we previously developed a collision warning algorithm for intersections based on a connected vehicle environment. To validate the reliability of this multi-fusion warning method for unsignalized intersections and road sections, we designed simulations based on Matlab and Simulink, using multiple real-world accident cases from Guilin, Guangxi. The simulations consist of two parts: case studies and comparative algorithm simulations. The case studies test the effectiveness and practicality of the warning model, while the comparative algorithm simulations provide a comprehensive statistical assessment of the model's reliability.

[0245] (1) Case simulation analysis

[0246] In order to verify the effectiveness of the early warning system based on the present invention, this embodiment selects a real accident case for simulation analysis. In this case: the vehicle's initial speed is 52.72km / h, the braking distance is 17.14m, and the average deceleration (MFDD) is 7.35m / s 2 The driver and vehicle braking reaction time is 1.240s. In the simulation, the vehicle is set to travel at a constant speed of 12.25m / s from the starting coordinates (16.5, -40.0). At the same time, the pedestrian walks at a normal speed from the starting coordinates (10.67, 3.99). At this moment, the velocity components are Vx = 1.45m / s and Vy = 0.29m / s.

[0247] Figure 11 and Figure 12 The simulation shows that when the vehicle is at the time T=0s, the distribution of the future collision time and distance is calculated by the multi-fusion warning method, and the collision distance and time show an obvious normal distribution trend. Figure 11 In the data, most of the sampled collision times are close to 4.2 seconds, indicating that the collision times are mainly concentrated in this time period; Figure 12 The data shows that the collision frequency between vehicles and pedestrians peaks between 63 and 64 meters, indicating a significant increase in collision risk within this distance range. In summary, the mean of the distribution indicates a future collision time of 4.2 seconds or a future collision distance of 63.5 meters, indicating the presence of a potential high-risk collision point.

[0248] In the subsequent simulation time, the experiment used the multi-fusion warning method to calculate the joint collision probability between vehicles and pedestrians in real time, and recorded the simulation scenarios and changes in the joint collision probability at key time nodes (T = 1s, T = 1.45s, T = 4.2s), such as Figure 13-15 shown.

[0249] like Figure 13 As shown in FIG, at time T=1s, the vehicle just starts to detect the pedestrian and the joint collision probability begins to gradually increase. Figure 13 -a and Figure 13 -b shows the real scene and simulation scene at this time, Figure 13 -c shows the curve of the combined collision probability. At this point, the combined collision probability has not exceeded the warning threshold of 0.55, and the system is in monitoring mode.

[0250] like Figure 14 As shown, at the moment T = 1.45s, Figure 14 -a and 14-b present the real scene and the simulation scene at this time; Figure 14 -c shows the change in the combined collision probability without warning. As can be seen from the figure, the combined collision probability rapidly climbs to nearly 98%, far exceeding the warning threshold. This indicates that if the vehicle does not brake in time, the probability of a collision with a pedestrian is extremely high.

[0251] like Figure 15 As shown in Figure 2, at the moment of T=4.2s, it indicates that the vehicle and pedestrian have collided in a real accident. Figure 14 -a and 14-b. If the early warning is given, Figure 15 -c shows the change in the combined collision probability after the warning. Since the collision reaction time between the driver and the vehicle is 1.240s, the combined collision probability rises rapidly and reaches a peak during this time, but drops significantly and returns to zero due to the driver's timely braking. Figure 15-d, the vehicle and pedestrian successfully avoided collision under the warning. The simulation results verified the effectiveness of the warning system in high-risk scenarios and demonstrated the practicality of the unsignaled intersection collision warning system based on the Internet of Vehicles in practical applications.

[0252] ② Comparative algorithm verification

[0253] Vulnerable traffic participants, such as pedestrians, often exhibit erratic behavior on the road, particularly at intersections, where they are highly unpredictable. Therefore, vehicles must maintain a safe distance or provide sufficient reaction time to ensure safety. The braking distance safety distance model and the TTC model, both traditional algorithms, are applicable in various traffic scenarios and play a significant role in predicting and avoiding collisions. Therefore, these two models were selected as comparison algorithms in this example and used in conjunction with the present invention for simulation experiments.

[0254] This example uses real traffic accident data from various types of unsignalized intersections in Guilin, Guangxi, as simulation scenarios. These include intersections such as crosses, T-junctions, and staggered roads, to comprehensively evaluate the collision warning accuracy of the three models. Furthermore, the vehicle speed range was set to 30-60 km / h based on traffic safety regulations and actual accident data from unsignalized intersections on national and provincial highways. To ensure the scientific nature of the experiments and the rationality of the data, 200 sets of experiments were designed for pedestrians, as shown in Table 1.

[0255] Table 1 Intersection experimental results of different algorithms

[0256] Number of tests Number of warnings required False alarm rate False alarm rate Accuracy The present invention 200 150 4.8％ 7.4％ 94.5％ TTC 200 150 16.2％ 23.1％ 82％ Distance algorithm 200 150 17％ 18.1％ 86.5％

[0257] Table 1 shows the test results of different algorithms when motor vehicles and pedestrians are involved in traffic. As shown in Table 1, the method of the present invention has a missed alarm rate of 4.8%, which is much lower than the 16.2% of the TTC model and the 17% of the braking distance model; the false alarm rate is 7.4%, which is significantly lower than the 23.1% of the TTC model and the 18.1% of the braking distance model; and the accuracy rate is as high as 94.5%, which is better than the 82% of the TTC model and the 86.5% of the braking distance model. The results show that the method of the present invention outperforms the TTC model and the braking distance-based model in terms of missed alarm rate, false alarm rate, and accuracy, verifying the reliability of the multi-fusion warning method. In addition, the present invention can accurately predict the collision probability between vehicles and pedestrians, and has the advantages of numerical risk reflection, reasonable prejudgment, and flexible assessment.

[0258] This invention focuses on the conflict between motor vehicles and pedestrians at unsignalized intersections. Using a Monte Carlo algorithm as its core, it integrates the SocialGAN model and a pedestrian trajectory prediction model based on a Multi-HeadAttention mechanism, while also introducing an elliptical buffer domain algorithm. Traditional algorithms, such as the TTC model, ignore pedestrian lateral movement and dynamic changes, leading to false and missed warnings in complex scenarios. Braking distance models also struggle with inadequate warning adjustments and limited accuracy when handling scenarios with large speed variations. In contrast, this invention primarily benefits from the elliptical buffer domain collision point identification method and the application of V2X technology. The former improves collision prediction accuracy, while the latter enhances information acquisition capabilities in complex environments. Furthermore, the SocialGAN and Multi-HeadAttention pedestrian trajectory prediction models enhance the system's ability to predict future pedestrian trajectories, reducing the occurrence of missed and false alarms. Comparative experiments, combining real-world data with traditional algorithms, effectively validated the superiority of this algorithm in predicting and warning potential collision risks.

[0259] The above description of the present invention and its embodiments is non-limiting and the actual embodiments are not limited thereto. In short, if a person skilled in the art is inspired by the above description and designs an implementation method and embodiment similar to the technical solution without departing from the purpose of the present invention, they shall fall within the scope of protection of the present invention.

Claims

1. A multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment, characterized by: The steps include: 1) Collecting and sharing information about motor vehicles and pedestrians; the information includes movement parameters and real-time locations of motor vehicles and pedestrians; 2) Input the information of motor vehicles and pedestrians into the vehicle trajectory model and pedestrian trajectory model respectively to calculate the future movement trajectory of the vehicle and pedestrian; 3) Based on the future motion trajectories of vehicles and pedestrians, an elliptical buffer algorithm is used to preliminarily screen pedestrians and identify potential collision targets within the buffer zone. 4) The historical trajectory of the identified potential collision target is input into the noisy pedestrian trajectory model to generate a large number of similar predicted trajectories; 5) Based on the predicted trajectory generated in step 4), using the elliptical buffer algorithm, adjusting the parameter K of the elliptical buffer, calculating and storing the potential collision point, potential collision time, and potential collision distance between the vehicle and the potential collision target; 6) Based on the potential collision point, potential collision time, and potential collision distance obtained in step 5), a Monte Carlo algorithm is used to calculate the collision expectation value and collision probability. Combined with the time (TTM) from the driver's awareness of the danger to the vehicle's complete stop, a joint collision probability is calculated based on a joint probability formula; 7) Compare the combined collision probability obtained in step 6) with a preset warning threshold. If the combined collision probability exceeds the preset threshold, the collision warning system is immediately triggered to alert the driver.

2. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 1 is characterized in that: The motion parameters of the motor vehicle in step S1 include driving speed, driving direction, and angular velocity; the motion parameters of the pedestrian include driving speed, driving direction, and steering angle; The real-time positions of the motor vehicles and pedestrians include GPS coordinates or roadside relative coordinates.

3. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 1 is characterized in that: The vehicle trajectory model in step S2 is a state vector P that can update the vehicle's driving direction and position information in real time based on a two-dimensional plane kinematic model. V (t), used to predict the future trajectory of the vehicle in the plane; The state vector P V (t) See formula (1): P V (t)=[x(t),y(t),v x (t),v y (t)] T (1) In formula (1), x(t) and y(t) represent the position coordinates of the vehicle on the two-dimensional plane at time t. The formulas for x(t) and y(t) are shown in the following formula (2): x(t-1) and y(t-1) represent the position of the vehicle at the previous moment t-1, and the velocity component of the vehicle at moment t is represented by v x (t) and v y (t) indicates that P V Corresponding to the vehicle's speed in the x and y directions respectively.

4. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 1 is characterized in that: The pedestrian trajectory model in step S2 is constructed based on the Social GAN ​​model, which incorporates a multi-head attention mechanism into the generator input and introduces a diversity loss function.

5. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 4 is characterized in that: The objective function of the adversarial neural network of the pedestrian trajectory model is shown in the following formula (3): Its core is to generate realistic future trajectories through adversarial training between the generator G and the discriminator D. The generator G tries to generate a trajectory G(z) that can "fool" the discriminator D, while the discriminator D tries to distinguish between the real trajectory x and the generated trajectory G(z); In the multi-head attention mechanism, the relationship between each time step is calculated as shown in equations (4) to (6): Z=MultiHead(Q,K,V) (5) G(z)=Generator(z,Z) (6) Where Q, K, and V are query, key, and value matrices respectively, T is the transpose, and d k is the dimension of the key; Z is the output of the multi-head self-attention mechanism, G(z) is the trajectory generator, and z represents the latent variable; Autonomous attention loss L introduced in pedestrian trajectory model att and diversity loss L var , see the following equations (7) and (8): In the above formula (7), A gen Represents the self-attention matrix generated by the model, A real represents the true self-attention matrix of the model; in formula (8), Y pred (k) Represents the k-th predicted trajectory. pred (k ' ) Represents the k'th predicted trajectory, which is another trajectory generated by the model and is the same as Y pred (k) There are differences between The total loss function of the adversarial neural network is shown in the following formula (9): L total =L GAN +λ1L att +λ2L var (9) Where λ1 and λ2 are hyperparameters, L GAN is the adversarial loss function.

6. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 5 is characterized in that: In step S4, the prediction trajectory formula generated by introducing noise into the pedestrian trajectory model is shown in the following formula (10): T i k =G(H k ,z)+N(0,σ 2 ) (10) Where, T i k represents the predicted trajectory of the i-th pedestrian of the k-th potential collision target, H k represents the historical trajectory of the input, z is a random noise vector, N(0,σ 2 ) is a Gaussian noise term used to perturb the trajectory at each generation to increase diversity.

7. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 1 is characterized in that: The basic formula of the buffer domain algorithm in steps 3) and 5) is shown in the following formula (11): Where x c and y c are the center points of the ellipse, a and b are the major and minor axes of the ellipse, θ is the rotation angle of the ellipse; x and y are the coordinates of the pedestrian or vehicle target point, respectively; The major axis and minor axis of the algorithm are defined as follows: d b =K·v vul ·t rea (14) Where L and W are the length and width of the vehicle respectively, d b Buffer distance used to extend the physical boundaries of the vehicle; v vul is the speed of the pedestrian, t rea is the driver's reaction time, and K is the adjustment coefficient.

8. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 7 is characterized in that: in, In step 3), pedestrians are preliminarily screened using the buffer domain algorithm using the following formula (15): Specifically, the four vertices of the pedestrian rectangle are substituted into the rotation ellipse equation of formula (15) to determine whether it is inside the ellipse; In step 5), the potential collision point, potential collision time, and potential collision distance between the vehicle and the potential collision target are calculated using the following equations (16) and (17): y j =m j x+e j ,j=1,2,3,4 (16) In the above formula, j represents the four sides of the rectangle, where the upper and lower, left and right boundaries are each a straight line; m j is the slope of the jth edge, e j is the intercept of the jth side; Specifically, the four vertices and edges of the pedestrian rectangle are substituted into the geometric equation of the rotation ellipse in equation (16) (17).

9. The multi-fusion warning method for vehicle-pedestrian collision risk at a non-signaled intersection in a V2X environment according to claim 1 is characterized in that: The joint probability formula in step 6) is shown in the following formula (24): P(t≤TTM,TTC∈R)=P(t≤TTM|TTC∈R)P(TTC∈R) (24) The TTC in the above formula represents the time required for a traffic participant to collide. The joint probability P(t≤TTM,TTC∈R) reflects the situation in which a collision occurs (TTC∈R) and the collision time t is less than TTM in all samples. The conditional probability P(t≤TTM / TTC∈R) is used to describe the possibility that the collision time t is less than TTM under the condition that the collision has already occurred (that is, TTC∈R); The collision probability P(TTC∈R) represents the proportion of TTCs in all samples that fall into the risk interval R; The expected value E(d) of the collision distance distribution and the collision probability P(TTC∈R) calculated by Monte Carlo are shown in the following equations (20) to (22): N total =m*h (20) In the above formula (20), h represents the number of potential targets, m represents the number of future trajectories generated by the same potential target; in formula (21), d k represents the collision distance of the kth potential target; N in formula (22) total is the total number of sampling times of the Monte Carlo algorithm, and II is the indicator function used to indicate whether there is a collision; In formula (22), C i k is a binary variable representing the collision judgment result, as shown in formulas (18) and (19): Where, T i k represents the i-th trajectory of the k-th potential target, m and h are natural integers, and N represents the number of time steps in the trajectory; C i k The elliptical buffer domain formula (16) is used to determine whether the kth potential collision target collides with the vehicle trajectory in the i-th predicted trajectory; The calculation of TTM is shown in the following formula (23): TTM=t rea +t brak (23) Where h represents the number of potential targets, m represents the number of future trajectories generated by the same potential target, and N total is the total number of Monte Carlo sampling, TTM is the driver and vehicle reaction time t rea and braking time t brak The sum of .

10. An electronic device comprising a memory, a processor, and a program stored in the memory, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 9 are implemented.

Images