A method for non-motor vehicle trajectory prediction

Abstract

The application relates to a non-motor vehicle trajectory prediction method, which comprises the following steps: calculating a feasible domain of a terminal point of a non-motor vehicle to be detected based on to-be-predicted scene information, and obtaining a feasible terminal point set through discrete sampling; adopting a deep learning model which is constructed and trained in advance, and generating an alternative trajectory set of the non-motor vehicle to be detected according to the to-be-predicted scene information and the feasible terminal point set; adopting a behavior analysis model of the non-motor vehicle which is constructed in advance, evaluating the alternative trajectory set, and selecting a maximum utility trajectory; and the behavior analysis model is used for calculating trajectory utility, and the trajectory utility is quantified as a sum of expected trajectory risk and trajectory efficiency heterogeneity. Compared with the prior art, the application can adapt to the flexible and changeable behavior characteristics of the non-motor vehicle, has certain advantages in eliminating cumulative errors in iterative prediction and trajectory prediction for sudden behaviors of the non-motor vehicle, and greatly improves the accuracy and reliability of the non-motor vehicle trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of traffic flow trajectory prediction, and in particular to a method for predicting the trajectory of non-motorized vehicles. Background Technology

[0002] Autonomous driving has entered the open road testing phase. However, open roads contain a large amount of shared space for motorized and non-motorized vehicles. The unclear right-of-way and lack of lane rules in these shared spaces, such as intersections, lead to complex interactions among various traffic participants, posing a challenge for autonomous vehicles navigating open roads. To address these challenges, predicting the behavior of other traffic participants is an effective method for autonomous vehicles to achieve safe interaction and avoid collisions in shared spaces. However, autonomous driving road test reports show that understanding complex scenarios and handling emergency situations remain critical issues for autonomous vehicles. Non-motorized vehicles, due to their high mobility, unpredictable riding, and sudden behavioral changes, present one of the main challenges in trajectory prediction.

[0003] Currently, research on non-motorized vehicle trajectory prediction can be divided into two categories: prediction methods based on physical models and prediction methods based on data. Physical models include kinematic models and dynamic models, and their main prediction principle is to extrapolate trends based on historical motion states. Data-driven prediction methods mainly rely on deep learning, using historical trajectories and future trajectories as inputs and outputs for training and learning, and then performing iterative predictions.

[0004] However, current iterative prediction methods focus more on extrapolating historical trends, but unavoidable cumulative errors exist during the iterative prediction process, making them more suitable for behaviors with strong regularity. Non-motorized vehicles, due to their flexible and varied behavior, exhibit significant trajectories, and may even show abrupt changes. Therefore, current trajectory prediction methods cannot fully meet the prediction needs of non-motorized vehicles. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing non-motorized vehicle trajectory prediction technologies, such as the cumulative error caused by iterative prediction and the inability to effectively predict the sudden changes in non-motorized vehicle behavior, and to provide a non-motorized vehicle trajectory prediction method.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for predicting the trajectory of non-motorized vehicles, applicable to mixed traffic conditions at intersections, includes the following steps:

[0008] S1: Obtain the scene information to be predicted, calculate the feasible domain of the destination of the non-motorized vehicle to be tested, and obtain the set of feasible destinations through discrete sampling;

[0009] S2: Using a pre-built and trained deep learning model, a set of alternative trajectories for the non-motorized vehicle to be tested is generated based on the information of the scene to be predicted and the set of feasible endpoints.

[0010] S3: Using a pre-constructed non-motorized vehicle behavior analysis model, the set of candidate trajectories is evaluated, and the trajectory with the maximum utility is selected; the behavior analysis model is used to calculate the trajectory utility, which is quantified as the heterogeneity sum of expected trajectory risk and trajectory efficiency.

[0011] Furthermore, the scenario information to be predicted includes information on the non-motorized vehicle to be tested, information on other traffic participants, and environmental information; the prediction process for all feasible destinations of the non-motorized vehicle includes:

[0012] The maximum and minimum displacements of the non-motorized vehicle under test in the forward direction and lateral direction are calculated based on the information of the scenario to be predicted, and the feasible region of the destination is generated. The occupied areas of other traffic participants are removed from the feasible region to obtain the feasible region of the destination of the non-motorized vehicle under test.

[0013] Furthermore, the calculation expressions for the maximum and minimum displacements of the non-motorized vehicle under test in the forward direction and laterally are as follows:

[0014]

[0015]

[0016] In the formula, x max x min y max and y min These represent the farthest and shortest distances that the non-motorized vehicle under test can reach along the x-axis and y-axis within the predicted time, respectively; x current and y current These are the current x and y coordinates of the non-motorized vehicle being measured; and These are the current speeds of the non-motorized vehicle under test in the x and y directions, respectively. and These are the maximum accelerations of the non-motorized vehicle under test in the positive and negative directions of the x and y axes, respectively; the calculation expression for removing the occupied area of ​​other traffic participants in this feasible region is:

[0017]

[0018] In the formula, x occ and y occ Ω represents the position of an object within the feasible region of the non-motorized vehicle to be tested; Ω represents the position of an object within the feasible region of (x occ ,y occ The spatial shape of the centroid; An empty set indicates that the region is not considered within the feasible region.

[0019] Furthermore, the discrete sampling specifically refers to spatial discretization sampling based on a rhombus-shaped occupant.

[0020] Furthermore, the deep learning model is used to generate the midpoint information of the trajectory based on the input trajectory start-point information and trajectory end-point information; the specific process of generating the candidate trajectory set of the non-motorized vehicle to be tested is as follows:

[0021] Based on the set of starting points and feasible endpoints of the non-motorized vehicle to be tested, a trajectory generation algorithm is adopted, and the trajectory endpoints are continuously generated through the deep learning model to obtain the trajectory of the non-motorized vehicle; by traversing each endpoint in the set of feasible endpoints, the set of candidate trajectories is obtained.

[0022] Furthermore, the calculation process of the trajectory generation algorithm includes the following steps:

[0023] S201: Select the i-th endpoint in the set of feasible endpoints as the current trajectory generation endpoint, and set the iteration number K;

[0024] S202: Generate complete start and end point feature vectors based on start point information and known end point location;

[0025] S203: Load the start and end point feature vectors into the trained deep learning model to generate the absolute position coordinates of the intermediate trajectory points;

[0026] S204: Convert the absolute position coordinates of the intermediate trajectory points into proportional coordinates;

[0027] S205: Based on all known trajectory point information, select the endpoint position, return to step S202, and execute K = K-1;

[0028] S206: Determine whether K=0 is true. If it is true, proceed to step S207. If it is false, repeat step S205.

[0029] S207: Set i = i + 1, and repeat steps S201 to S206 until the entire preset trajectory set generation task is completed.

[0030] Furthermore, the complete start-end feature vector includes parameters of the start and end positions of the non-motorized vehicle and the interactive object under test, including coordinates, velocity, acceleration, curvature, and vehicle type;

[0031] The expressions for calculating the velocity and acceleration at the endpoint of the non-motorized vehicle and the interactive object are as follows:

[0032]

[0033]

[0034] In the formula, P, V, A, and s represent position coordinates, velocity, acceleration, and vehicle type, respectively; where P = {x, y}, V = {v_x, v_y}, and A = {a_x, a_y} represent the position, velocity, and acceleration of the research object on the x and y coordinate axes, respectively; s = {0, 1, 2} represent motor vehicles, electric bicycles, and regular bicycles, respectively; the superscript host and int represent the research subject and the interaction object, respectively; the subscripts start, end, and mid represent the start and end points and intermediate trajectory points, respectively; t represents the prediction time; the interaction object is the other traffic participant whose Euclidean distance to the non-motor vehicle under test is closest at the current moment.

[0035] Furthermore, the calculation expression for the trajectory utility is as follows:

[0036]

[0037]

[0038]

[0039] In the formula, B(A) i ) is the utility value of the i-th trajectory, A i The coordinate matrix representing the i-th trajectory. Represents the position coordinates of the j-th trajectory point of the i-th trajectory; R and L represent path risk and path efficiency, respectively. Here, R is negative, indicating that it is in the opposite direction to L. The cycling preference coefficient represents the heterogeneous pursuit of risk and efficiency by individuals.

[0040] Furthermore, the calculation expression for the path risk is as follows:

[0041]

[0042] E = |E r |+|E v |

[0043] In the formula, R represents path risk; d represents the path length; E j The risk value represents the location of the j-th trajectory point, and there are n trajectory points in total; E represents the risk value of a certain point, which is a scalar; E r and E v These represent the potential energy field and the dynamic field, respectively, where the potential energy field and the dynamic field represent the risk distribution of static and dynamic objects, respectively.

[0044] The dynamic field E vThe sequence prediction position calculation based on dynamic interactive objects is expressed as follows:

[0045] P dy =V dy_start T+P dy_start

[0046] In the formula, P dy V represents the position matrix of the dynamic interactive object at each step within the prediction time; dy_start P represents its initial velocity; T represents the time step matrix; P represents its initial velocity. dy_start Indicates the initial position coordinates.

[0047] Furthermore, the expression for calculating the path efficiency is as follows:

[0048]

[0049] In the formula, L represents the path efficiency; l j Let be the forward displacement of the j-th step, and n be the total number of trajectory points.

[0050] Compared with the prior art, the present invention has the following advantages:

[0051] (1) This invention proposes a method for predicting the trajectory of non-motorized vehicles going straight through intersections. Based on a set of feasible endpoints, a set of feasible trajectories is generated using a deep learning model. By constructing a behavior analysis model, the trajectory with the maximum utility is selected, thus achieving an effective integration of domain knowledge and data-driven methods. The model can leverage the ability of data-driven methods to extract the overall complex features of non-motorized vehicle execution trajectories, achieving accurate output of the overall trajectory. At the same time, due to the integration of domain knowledge, the rationality of the prediction results at the behavior decision layer can be guaranteed, achieving an accurate description of the sudden changes in non-motorized vehicle behavior, thereby compensating for the unreasonable output of data-driven methods when facing sparse samples.

[0052] (2) This invention proposes a new trajectory encoding and decoding method for learning and generating non-motorized vehicle trajectories. Its advantages are: by taking the overall trajectory as the modeling object, it can accurately extract the trajectory trend features, avoid the cumulative error of the traditional iterative prediction method, and ensure the fitting effect of the generated trajectory set on the real trajectory. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating a non-motorized vehicle trajectory prediction method provided in an embodiment of the present invention;

[0054] Figure 2 This is a trajectory generation result diagram provided in an embodiment of the present invention;

[0055] Figure 3This is a predicted trajectory result diagram provided in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0057] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0058] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0059] This embodiment provides a method for predicting the trajectory of non-motorized vehicles, including the following steps:

[0060] S1: Obtain the scene information to be predicted, calculate the feasible domain of the destination of the non-motorized vehicle to be tested, and obtain the set of feasible destinations through discrete sampling;

[0061] S2: Using a pre-built and trained deep learning model, a set of alternative trajectories for the non-motorized vehicle to be tested is generated based on the information of the scene to be predicted and the set of feasible endpoints.

[0062] S3: A pre-built behavioral analysis model for non-motorized vehicles is used to evaluate the set of candidate trajectories and select the trajectory with the maximum utility. The behavioral analysis model is used to calculate the trajectory utility, which is quantified as the sum of the heterogeneity of expected trajectory risk and trajectory efficiency.

[0063] For step S1, the scenario information to be predicted includes information on the non-motorized vehicle to be tested, information on other traffic participants, and environmental information; the feasible region prediction process for the non-motorized vehicle's destination includes:

[0064] The maximum and minimum displacements of the non-motorized vehicle under test in the forward and lateral directions are calculated based on the information of the scenario to be predicted, and the feasible region of the destination is generated. The occupied areas of other traffic participants are removed from the feasible region to obtain the feasible region of the destination of the non-motorized vehicle under test.

[0065] The calculation expressions for the maximum and minimum displacements of the non-motorized vehicle under test in the forward and lateral directions are as follows:

[0066]

[0067]

[0068] In the formula, x max x min y max and y min These represent the farthest and shortest distances that the non-motorized vehicle under test can reach along the x-axis and y-axis within the predicted time, respectively; x current and y current These are the current x and y coordinates of the non-motorized vehicle being measured; and These are the current speeds of the non-motorized vehicle under test in the x and y directions, respectively. and These are the maximum accelerations of the non-motorized vehicle under test in the positive and negative directions of the x and y axes, respectively; the calculation expression for removing the occupied areas of other traffic participants within this feasible region is:

[0069]

[0070] In the formula, x occ and y occ Ω represents the position of an object within the feasible region of the non-motorized vehicle to be tested; Ω represents the position of an object within the feasible region of (x occ ,y occ The spatial shape of the centroid; An empty set indicates that the region is not considered within the feasible region.

[0071] Discrete sampling specifically involves spatial discretization sampling based on a rhombus-shaped occupant.

[0072] For step S2, the deep learning model is used to generate the midpoint information of the trajectory based on the input trajectory start-point information and trajectory end-point information; the specific process of generating the candidate trajectory set for the non-motorized vehicle to be tested is as follows:

[0073] Based on the set of starting points and feasible endpoints of the non-motorized vehicle to be tested, a trajectory generation algorithm is adopted, and the trajectory endpoints are continuously generated through a deep learning model to obtain the trajectory of the non-motorized vehicle; by traversing each endpoint in the set of feasible endpoints, a set of candidate trajectories is obtained.

[0074] The computation process of the trajectory generation algorithm includes the following steps:

[0075] S201: Select the i-th endpoint in the set of feasible endpoints as the current trajectory generation endpoint, and set the iteration number K;

[0076] S202: Generate complete start and end point feature vectors based on start point information and known end point location;

[0077] S203: Load the start and end point feature vectors into the trained deep learning model to generate the absolute position coordinates of the intermediate trajectory points;

[0078] S204: Convert the absolute position coordinates of intermediate trajectory points into proportional coordinates;

[0079] S205: Based on all known trajectory point information, select the endpoint position, return to step S202, and execute K = K-1;

[0080] S206: Determine whether K=0 is true. If it is true, proceed to step S207. If it is false, repeat step S205.

[0081] S207: Set i = i + 1, and repeat steps S201 to S206 until the entire preset trajectory set generation task is completed.

[0082] The complete start-end feature vector includes parameters of the start and end positions of the non-motorized vehicle and the interactive object under test. These parameters include coordinates, velocity, acceleration, curvature, and vehicle type.

[0083] The expressions for calculating the velocity and acceleration at the endpoint of the non-motorized vehicle and the interactive object are as follows:

[0084]

[0085]

[0086] In the formula, P, V, A, and s represent position coordinates, velocity, acceleration, and vehicle type, respectively; where P = {x, y}, V = {v_x, v_y}, and A = {a_x, a_y} represent the position, velocity, and acceleration of the research object on the x and y coordinate axes, respectively; s = {0, 1, 2} represent motor vehicles, electric bicycles, and regular bicycles, respectively; the superscript host and int represent the research subject and the interaction object, respectively; the subscripts start, end, and mid represent the start and end points and intermediate trajectory points, respectively; t represents the prediction time; and the interaction object is the other traffic participant whose Euclidean distance to the non-motor vehicle under test is closest at the current moment.

[0087] For step S3, the calculation expression for trajectory utility is:

[0088]

[0089]

[0090]

[0091] In the formula, B(A) i) is the utility value of the i-th trajectory, A i The coordinate matrix representing the i-th trajectory. Represents the position coordinates of the j-th trajectory point of the i-th trajectory; R and L represent path risk and path efficiency, respectively. Here, R is negative, indicating that it is in the opposite direction to L. The cycling preference coefficient represents the heterogeneous pursuit of risk and efficiency by individuals.

[0092] The formula for calculating path risk is:

[0093]

[0094] E = |E r |+|E v |

[0095] In the formula, R represents path risk; d represents the path length; E j The risk value represents the location of the j-th trajectory point, and there are n trajectory points in total; E represents the risk value of a certain point, which is a scalar; E r and E v These represent the potential energy field and the dynamic field, respectively, where the potential energy field and the dynamic field represent the risk distribution of static and dynamic objects, respectively.

[0096] Dynamic field E v The sequence prediction position calculation based on dynamic interactive objects is expressed as follows:

[0097] P dy =V dy_start T+P dy_start

[0098] In the formula, P dy V represents the position matrix of the dynamic interactive object at each step within the prediction time; dy_start P represents its initial velocity; T represents the time step matrix; P represents its initial velocity. dy_start Indicates the initial position coordinates.

[0099] The expression for calculating path efficiency is:

[0100]

[0101] In the formula, L represents the path efficiency; l j Let be the forward displacement of the j-th step, and n be the total number of trajectory points.

[0102] Example 1

[0103] The overall process of this embodiment is as follows: Figure 1As shown, firstly, the feasible domain of the destination of the non-motorized vehicle is generated based on kinematics and safety constraints, and a finite complete set of destinations is obtained by discrete sampling. Then, a deep learning model is used to learn and generate a complete set of alternative trajectories for the non-motorized vehicle. Finally, a behavioral analysis model of the non-motorized vehicle is constructed based on the balance between expected trajectory risk and efficiency to evaluate and select alternative trajectories, and then output the predicted trajectory.

[0104] In this embodiment, verification data was collected at the intersection of Xianxia Road and Jianhe Road in Shanghai. The intersection is 50m long from east to west, with lane markings separating non-vehicle traffic on both upstream and downstream sections, and the non-motorized vehicle lane is 3.5m wide. The video acquisition device was installed on a high-rise building next to the intersection to obtain a high-altitude perspective. The video was collected during the evening rush hour (4:30 PM - 5:30 PM). After acquiring the video data, this paper used video trajectory extraction software to extract the trajectories of non-motorized vehicles and other traffic participants, with an extraction step size of 0.12s. A total of 680 non-motorized vehicle trajectories were obtained, with tags including trajectory coordinates, speed, acceleration, curvature, and vehicle type. In this embodiment, the prediction time was 1.92s (17 steps).

[0105] The specific implementation of this embodiment is described through the following steps:

[0106] Step 1: Generate all feasible destinations for non-motorized vehicles within the next 1.92 seconds (17 steps).

[0107] First, prediction data is input based on the prediction scenario, including information on non-motorized vehicles, other traffic participants, and the environment. Parameters include: position coordinates, velocity, acceleration, curvature, vehicle type, distance between the target and the non-motorized vehicle separation line, and distance between the target and the exit lane. Then, based on the non-motorized vehicle's direction of motion, it is divided into forward and lateral directions. The maximum and minimum displacements of the non-motorized vehicle in the forward and lateral directions are calculated based on the 80th percentile of the non-motorized vehicle's acceleration statistics, generating the feasible region of the destination. The calculation formula is as follows:

[0108]

[0109]

[0110] In the formula, x max x min y max and y min These represent the farthest and shortest distances that non-motorized vehicles can reach along the x-axis and y-axis within the predicted time, respectively; x current and y current These are the current x and y coordinates of the research subject, respectively; and These are the current velocities of the research subject in the x and y directions, respectively. and These represent the maximum accelerations of non-motorized vehicles in the positive and negative directions of the x and y axes, respectively.

[0111] Since the entities of other traffic participants do not have endpoint feasibility, non-motorized vehicles and motorized vehicles are removed by using the diamond-shaped area as the occupancy area. The calculation formula is as follows:

[0112]

[0113] In the formula, x occ and y occ Indicates the location of an object within the feasible end area of ​​a non-motorized vehicle; Ω represents the position of an object within (x... occ ,y occ The shape of the space is defined by the centroid, where non-motorized vehicles and motorized vehicles each take rhomboid regions of different sizes. An empty set indicates that the region is not included in the feasible endpoint region.

[0114] The feasible region for non-motorized vehicles has an infinite number of feasible endpoints. To control the number of feasible endpoints, the feasible region of the endpoints is discretized to obtain a finite number of feasible endpoints. The sampling method uses a diamond-shaped occupancy as a reference for spatial discretization sampling. The ratio of the major axis to the minor axis of the diamond-shaped sampling area can be flexibly set according to the prediction scenario to ensure the completeness of sampling in both the forward and lateral directions. The calculation formula is expressed as follows:

[0115] S=Ω(a,σa)

[0116] In the formula, S represents the sampling space; Ω(a,σa) represents the rhomboid region, where a is the major axis of the rhombus, describing the sampling length of the space in the direction of non-motorized vehicle movement. The value of a affects the prediction accuracy; σa is the minor axis of the rhombus, describing the sampling length of the lateral space in the movement of non-motorized vehicles, where 0 < σ < 1. The smaller the value, the denser the sampling of the lateral space. By adjusting the values ​​of σ and a, the density of spatial sampling is changed, achieving a balance between the algorithm's running speed and prediction accuracy. In this embodiment, a and σ are set to 0.4m and 0.5 respectively. In other embodiments, the parameters can be modified as needed.

[0117] Step 2: Learn and generate a set of candidate trajectories for non-motorized vehicles based on a convolutional neural network (CNN) model.

[0118] In this step, a CNN deep learning model is used, including model training and trajectory generation. The input data for model training is trajectory data, with parameters including coordinates, velocity, acceleration, curvature, vehicle type, and the parameters of the interactive object with the shortest Euclidean distance to the non-motorized vehicle. The output data includes the position, velocity, acceleration, and vehicle type of the midpoint between the non-motorized vehicle and the interactive object. The velocity and acceleration at the endpoint in the input data are unknowns during prediction, generated using kinematic formulas based on current information. The calculation formula is as follows:

[0119]

[0120]

[0121] In the formula, P, V, A, and s represent position coordinates, velocity, acceleration, and vehicle type, respectively; where P = {x, y}, V = {v_x, v_y}, A = {a_x, a_y}, representing the position, velocity, and acceleration of the research object on the x and y coordinate axes, respectively; s = {0, 1, 2}, representing motor vehicles, electric bicycles, and regular bicycles, respectively; the superscripts host and int represent the research subject and the interaction object, respectively; the subscripts start, end, and mid represent the start-end point and intermediate trajectory point, respectively. t represents the prediction time. The interaction object is the other traffic participant whose Euclidean distance to the non-motorized vehicle is closest at the current time.

[0122] Furthermore, the position information of the midpoint in the output data of the model training is replaced by proportional coordinates instead of absolute position coordinates, and the calculation formula is as follows:

[0123]

[0124] In the formula, P transform For the transformed coordinate matrix, 0 < |P transform | < 1; This indicates the absolute position coordinates of the output trajectory points.

[0125] In summary, the trajectory encoding algorithm for model training is summarized as follows:

[0126] Table 1 Trajectory Encoding Algorithm

[0127]

[0128]

[0129] In the formula, j is the trajectory number in the training sample sampling; k is the sampling level in the training sample sampling, which represents the number of trajectories between the output sample point and the input sample point during the sampling process; j maxIt is the maximum number of trajectories in the j-th trajectory. In this embodiment, after cyclic sampling and trajectory encoding, a total of 46301 training samples were obtained.

[0130] Trajectory generation is used to generate all possible alternative trajectories for non-motorized vehicles within the predicted time. The algorithm is as follows:

[0131] Table 6 Trajectory Generation Algorithm

[0132]

[0133] In the table, ① refers to the number of iterations K, which is the number of times the trajectory is encrypted using the trajectory generation model.

[0134] Step 3: Construct a behavioral analysis model for non-motorized vehicles, evaluate and select the trajectory with the highest utility.

[0135] This step includes non-motorized vehicle behavior analysis, trajectory utility calculation, and trajectory selection. The trajectory utility can be quantified as the sum of the heterogeneity of expected trajectory risk and trajectory efficiency, calculated using the following formula:

[0136]

[0137]

[0138]

[0139] In the formula, B(A) i ) is the utility value of the i-th trajectory, A i The coordinate matrix representing the i-th trajectory. Represents the position coordinates of the j-th trajectory point of the i-th trajectory; R and L represent path risk and efficiency, respectively. Here, R is negative, indicating that it is in the opposite direction to L. The cycling preference coefficient represents the heterogeneous pursuit of risk and efficiency by individuals.

[0140] Route risk is used to quantify the total risk faced by cyclists due to choosing different cycling routes within a predicted time period. The calculation formula is as follows:

[0141]

[0142] E = |E r |+|E v |

[0143] In the formula, R represents path risk; d represents the path length; E j The risk value represents the location of the j-th trajectory point, and there are n trajectory points in total; E represents the risk value of a certain point, which is a scalar; E r and E vThese represent the potential energy field and the dynamic field, respectively, where the potential energy field and the dynamic field represent the risk distribution of static and dynamic objects, respectively.

[0144] Furthermore, the path risk shown is based on the sequence prediction position of dynamic interactive objects, and the sequence prediction formula for these dynamic interactive objects is as follows:

[0145] P dy =V dy_start T+P dy_start

[0146] In the formula, P dy V represents the position matrix of the dynamic interactive object at each step within the prediction time; dy_start P represents its initial velocity; T represents the time step matrix; P represents its initial velocity. dy_start Indicates the initial position coordinates.

[0147] In this embodiment, path efficiency can be quantified as the sum of the positions of each step size in the forward direction:

[0148]

[0149] In the formula, L represents the path efficiency; l j Let be the forward displacement of the j-th step.

[0150] In addition, to ensure the accuracy of the model, kinematic constraints (i.e., velocity, acceleration, and deceleration) and safety constraints (i.e., maximum risk) are introduced.

[0151] Finally, based on the principle of maximizing utility, the predicted trajectory is output.

[0152] Step 4: Evaluation of Prediction Results

[0153] This embodiment verifies the trajectory prediction error based on test sets of all non-motorized vehicle behaviors and test sets of non-motorized vehicle mutation behaviors. Furthermore, it verifies the validity of the trajectory generation and selection results. The trajectory similarity evaluation criteria are Average distance error (ADE), Final distance error (FDE), and Mean Locality In-between Polylines (MLIP).

[0154] Table 3 Prediction error of test samples

[0155]

[0156] The prediction error of the fusion model is lower than that of the behavior analysis model, and both are lower than existing methods. Furthermore, the FDE of the fusion model is reduced by approximately 35% compared to the behavior analysis model, indicating that the cumulative error of predicting the trajectory endpoint is significantly reduced due to trajectory generation.

[0157] Table 4 shows the prediction error for large curvature behavior.

[0158]

[0159] Table 5 shows the prediction error for lateral offset behavior.

[0160]

[0161] like Figure 3 As shown, without loss of generality, two types of bicycle mutation behaviors were extracted as samples to analyze the generalization of mutation behaviors by the fusion model. The first type is high curvature behavior, defined as behavior where the average trajectory curvature within the top 5% of the overall samples during the prediction time. The other is lateral displacement behavior, defined as behavior where the lateral displacement of the trajectory exceeds 0.5 meters during the prediction time. The prediction results in the table show that the fusion model has high prediction accuracy for both types of mutation behaviors, with the fusion model showing higher prediction accuracy for lateral displacement behavior than for high curvature behavior.

[0162] Table 6 Trajectory Generation Results

[0163]

[0164] Analysis of the trajectory generation results shows that the accuracy of trajectory generation (ADE) for both the actual trajectory endpoint set and the feasible trajectory endpoint set is less than 0.1m (e.g., ...). Figure 2 As shown in the figure, it is demonstrated that the deep learning-based trajectory generation method proposed in this work can accurately generate intermediate trajectories with known starting and ending points.

[0165] Calculating the trajectory selection probability for the entire sample reveals that the selection probability of the behavioral analysis model is inversely proportional to the error between the selected trajectory and the ground reality. In other words, the majority of samples choose low-error trajectories, while the probability of choosing larger errors is very small. Approximately 80% of the samples selected the top five trajectories with the smallest errors, indicating that the modeling approach based on the balance between expected trajectory efficiency and risk can reliably describe bicycle decision-making behavior and predict accurate bicycle trajectories.

[0166] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A non-motor vehicle trajectory prediction method, characterized in that, Includes the following steps: S1: Based on the information of the scene to be predicted, calculate the feasible region of the destination of the non-motorized vehicle to be tested, and obtain the set of feasible destinations through discrete sampling; S2: Using a pre-built and trained deep learning model, a set of alternative trajectories for the non-motorized vehicle to be tested is generated based on the information of the scene to be predicted and the set of feasible endpoints. S3: Using a pre-constructed behavior analysis model for non-motorized vehicles, the set of candidate trajectories is evaluated, and the trajectory with the maximum utility is selected; the behavior analysis model is used to calculate the trajectory utility, which is quantified as the heterogeneity sum of expected trajectory risk and trajectory efficiency; The calculation expression for the trajectory utility is: wherein, is the utility value of the ith trajectory, represents the coordinate matrix of the ith trajectory, represents the position coordinates of the jth trajectory point of the ith trajectory; R and L represent the path risk and the path efficiency, respectively, R is negative here, indicating that R and L take opposite directions; represents the riding preference coefficient, indicating the heterogeneous pursuit of risk and efficiency by the individual; The expression for calculating the path risk is: In the formula, R For path risk; d The length of the path; Indicates the first j The risk value of each trajectory point location is totaled. n A trajectory point; E The risk value at a certain point is a scalar. and These represent the potential energy field and the dynamic field, respectively, where the potential energy field and the dynamic field represent the risk distribution of static and dynamic objects, respectively. The dynamic field The sequence prediction position calculation based on dynamic interactive objects is expressed as follows: wherein represents a position matrix of the dynamic interactive object at each step of the prediction time; represents its initial velocity; represents a time step matrix; represents initial position coordinates; The expression for calculating the path efficiency is: In the formula, L is the path efficiency; is the first j is the forward displacement of the step, is the total number of trajectory points.

2. The non-automotive vehicle trajectory prediction method of claim 1, wherein, The scenario information to be predicted includes information on the non-motorized vehicle to be tested, information on other traffic participants, and environmental information; the prediction process for all feasible destinations of the non-motorized vehicle includes: Based on the scenario information to be predicted, calculate the maximum and minimum displacements of the non-motorized vehicle in the forward direction and lateral direction to generate the feasible region of the destination; remove the occupied areas of other traffic participants in the feasible region to obtain the feasible regions of all destinations of the non-motorized vehicle to be tested.

3. The non-automotive vehicle trajectory prediction method of claim 2, wherein, The calculation expressions for the maximum and minimum displacements of the non-motorized vehicle under test in the forward and lateral directions are as follows: In the formula, , , and These represent the farthest and shortest distances that the non-motorized vehicle under test may reach along the x-axis and y-axis within the predicted time, respectively. and These are the current x and y coordinates of the non-motorized vehicle being measured; and These are the current speeds of the non-motorized vehicle under test in the x and y directions, respectively. , , and These are the maximum accelerations of the non-motorized vehicle under test in the positive and negative directions of the x and y axes, respectively; The calculation expression for removing the occupied areas of other traffic participants in the feasible region is as follows: wherein and represents the position of an object within the feasible region of the non-automotive vehicle to be tested; represents a spatial shape with as the center; is an empty set, indicating that the region is not counted in the feasible region.

4. The non-automotive vehicle trajectory prediction method of claim 1, wherein, The discrete sampling specifically involves spatial discretization sampling based on a rhombus-shaped occupier; the rhombus-shaped occupier is dynamically variable, and its expression is as follows: In the formula, For the surface of the sampling space; Represents a rhombus-shaped region, in which The major axis of the rhombus represents the sampling of the space describing the direction of travel of the non-motorized vehicle under test. The minor axis of the rhombus describes the sampling of the lateral space of the motion of the non-motorized vehicle under test, where... By adjusting and The value of can be changed to alter the density of spatial sampling, thus achieving a balance between the algorithm's running speed and prediction accuracy.

5. The non-automotive vehicle trajectory prediction method of claim 1, wherein, The deep learning model is used to generate the midpoint information of the trajectory based on the input trajectory start-point information and trajectory end-point information; the specific process of generating the candidate trajectory set of the non-motorized vehicle to be tested is as follows: Based on the set of starting points and feasible endpoints of the non-motorized vehicle to be tested, a trajectory generation algorithm is adopted, and the trajectory endpoints are continuously generated through the deep learning model to obtain the trajectory of the non-motorized vehicle; by traversing each endpoint in the set of feasible endpoints, the set of candidate trajectories is obtained.

6. The non-automotive vehicle trajectory prediction method of claim 5, wherein, The calculation process of the trajectory generation algorithm includes the following steps: S201: Select the first [number] from the set of feasible endpoints. i The current trajectory generation endpoint is set as one endpoint, and the iteration number K is set; S202: Generate complete start and end point feature vectors based on start point information and known end point location; S203: Load the start and end point feature vectors into the trained deep learning model to generate the absolute position coordinates of the intermediate trajectory points; S204: Convert the absolute position coordinates of the intermediate trajectory points into proportional coordinates; S205: Select the end point position according to all known trajectory point information, return to step S202, and execute ; S206: Determine whether K=0 is true. If it is true, proceed to step S207. If it is false, repeat step S205. S207: setting The steps S201 to S206 are repeatedly executed until the whole generation task of the preset trajectory set is completed.

7. The non-automotive vehicle trajectory prediction method of claim 5, wherein, The complete start-end feature vector includes parameters of the start and end positions of the non-motorized vehicle and the interactive object under test. These parameters include coordinates, velocity, acceleration, curvature, and vehicle type. The expressions for calculating the velocity and acceleration at the endpoint of the non-motorized vehicle and the interactive object are as follows: In the formula, P, V, A, and s represent position coordinates, velocity, acceleration, and vehicle type, respectively; where P={x,y}, V={v_x,v_y}, and A={a_x,a_y} represent the position, velocity, and acceleration of the research object on the x and y coordinate axes, respectively; s={0,1,2} represent motor vehicles, electric bicycles, and regular bicycles, respectively; the superscript host and int represent the research subject and the interaction object, respectively; and the subscripts start, end, and mid represent the start and end points and intermediate trajectory points, respectively. t The predicted time is indicated; the interaction object is the other traffic participant whose Euclidean distance to the non-motorized vehicle being tested is closest at the current moment.

Images