Non-motor vehicle adaptive trajectory prediction method

Abstract

The invention relates to a non-motor vehicle adaptive trajectory prediction method, which comprises the steps of generating an expected candidate trajectory set of a rider based on real-time scene information, evaluating each trajectory in the expected candidate trajectory set through a preset utility function, and selecting a theoretical optimal trajectory as a predicted value of the trajectory at the current moment; the actual observation value at the current moment is obtained, the deviation between the actual observation value and the corresponding predicted value is calculated, the personalized features of the rider are recognized and determined, and the personalized features comprise the aggressive level and the rational level of the rider; performing mode modeling according to the sequence of the personalized features to obtain a feature evolution rule so as to perform parameter adjustment on the trajectory prediction model; and carrying out personalized prediction on the individual rider by adopting the trajectory prediction model, and recalculating and updating the personalized characteristics of the rider according to the obtained predicted value to form closed-loop optimization. Compared with the prior art, the method has the advantage that the prediction capability of the model for heterogeneous riders in dynamic application is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of non-motor vehicle trajectory prediction, and in particular to a non-motor vehicle adaptive trajectory prediction method. Background Art

[0002] With the gradual implementation of autonomous driving technology in open road scenarios, how to achieve safe and efficient operation in complex mixed traffic environments has become an urgent technical problem to be solved. Autonomous vehicles rely on trajectory prediction algorithms to anticipate the future behaviors of interaction objects, and thus generate and optimize driving paths. Therefore, accurately predicting the future trajectories of other traffic participants is crucial for the reliability and safety of autonomous driving systems. In a traffic environment with mixed motor and non-motor vehicles, the frequent interaction between non-motor vehicles and motor vehicles makes non-motor vehicles the key interaction objects that autonomous driving systems need to focus on. However, non-motor vehicles have a high degree of mobility and are not restricted by lane rules, and their behavioral characteristics are complex and variable. This high degree of uncertainty poses a huge challenge to trajectory prediction and becomes a restricting factor for the reliable operation of autonomous driving systems in complex traffic environments.

[0003] Currently, trajectory prediction algorithms are mainly based on data-driven models, with "averaged feature learning" as the core feature, and rely on pre-trained models for iterative prediction. However, such methods have the following significant deficiencies: First, they cannot effectively capture and learn the personalized decision-making characteristics of cyclists, making it difficult to achieve targeted trajectory prediction; second, they have weak adaptability. Most model parameters are obtained through offline training and cannot be updated in real time to reflect the differences in decision-making strategies of heterogeneous cyclists, resulting in a significant decline in prediction accuracy in actual application scenarios.

[0004] Since non-motor vehicles are not restricted by behavioral rules and have a high degree of instability, cyclists often exhibit complex and variable behavioral characteristics. In this context, the immediate decisions of cyclists become the key factors affecting their future trajectories. Cyclists will make quick decisions based on the interaction environment and their own personalized needs and execute them immediately. Since the decision-making characteristics of cyclists are dynamic and heterogeneous, their interaction strategies also show random changes. Therefore, accurately analyzing and tracking the decision-making characteristics of cyclists is the core of achieving accurate trajectory prediction. However, the current static pre-trained trajectory prediction models are difficult to effectively handle dynamic interaction scenarios and cannot fully meet the real-time adaptability requirements of autonomous driving systems. Summary of the Invention

[0005] The purpose of the present invention is to provide a non-motor vehicle adaptive trajectory prediction method to overcome the deficiencies of the existing technology that static pre-trained trajectory prediction models are difficult to effectively handle dynamic interaction scenarios and cannot fully meet the real-time adaptability requirements of autonomous driving systems.

[0006] The object of the present invention can be achieved by the following technical solutions:

[0007] A non-motor vehicle adaptive trajectory prediction method, comprising the following steps:

[0008] S1: Based on real-time scenario information, generate a set of expected candidate trajectories for the rider, and evaluate each trajectory in the set of expected candidate trajectories through a preset utility function, and select the theoretically optimal trajectory as the predicted value of the trajectory at the current moment;

[0009] S2: Obtain the actual observation value at the current moment, calculate the deviation from the corresponding predicted value, identify and determine the personalized characteristics of the rider, and the personalized characteristics include the aggressive level and rational level of the rider;

[0010] S3: Perform pattern modeling according to the determined sequence of personalized characteristics to obtain the feature evolution law, so as to adjust the parameters of the pre-constructed trajectory prediction model;

[0011] S4: Use the adjusted trajectory prediction model to perform personalized prediction on the rider individual, and return to step S2 according to the obtained predicted value to update the personalized characteristics of the rider, forming a closed-loop optimization.

[0012] Further, the generation process of the set of expected candidate trajectories in step S1 includes the following steps:

[0013] S101: Based on real-time scenario information, determine the feasible end region of the rider according to the motion law;

[0014] S102: Discretely sample the feasible end region into a set of end points;

[0015] S103: According to the obtained end points, use the trajectory interpolation method to generate the corresponding candidate trajectories respectively, and generate a set of kinematically feasible expected candidate trajectories.

[0016] Further, the real-time scenario information includes the position, speed and acceleration information of the rider to be measured and other traffic participants, as well as environmental information.

[0017] Further, the feasible end region is determined by calculating the farthest reachable position and the nearest reachable position in the motion direction and the lateral direction.

[0018] Further, the expression of the utility function is:

[0019]

[0020] In the formula, U(S i ) is the utility value of candidate trajectory i, S i is the coordinate matrix of trajectory i, n is the total number of candidate trajectories, and represent the risk value and efficiency value of the candidate trajectory respectively, is the risk - efficiency preference, representing the aggressive level of the rider,

[0021] Furthermore, the calculation expression of the identification process of the aggressive level is:

[0022]

[0023] In the formula, is the aggressive level parameter of the theoretically optimal trajectory in the expected candidate trajectory set, S opt is the corresponding optimal trajectory, and represent the risk value and efficiency value of the trajectory S opt respectively, is the weight of the safety term in the rider's utility, and are the coordinate matrices of the trajectory prediction value and the trajectory observation value respectively, n is the total number of candidate trajectories, and represent the coordinates of the predicted trajectory point and the observed trajectory point in the time step j respectively, D(·) is the calculation method for measuring the similarity of the trajectory.

[0024] Furthermore, the online update process of the aggressive level includes:

[0025] S201: Update the rider's aggressive level sequence based on the actual observation value and the corresponding prediction value of the trajectory;

[0026] S202: Construct an aggressive level evolution model through pattern learning based on the existing sequence data of the aggressive level;

[0027] S203: Estimate the aggressive level parameter at the future moment based on the constructed aggressive level evolution model and use it to adjust the parameters of the trajectory prediction model.

[0028] Furthermore, the calculation expression of the identification process of the rational level is:

[0029]

[0030] In the formula, λ t represents the rational level characteristic parameter, N target represents the ranking of the selected trajectory in the candidate trajectory evaluation, N all represents the number of candidates.

[0031] Furthermore, the rational level is defined as: the ratio of the individual utility selected by the rider in the decision - making to the theoretical optimal individual utility;

[0032] The learning and estimation process of the rational level includes:

[0033] S301: Continuously collect the utility entropy sequence and the feature sequence of the rational level;

[0034] S302: Through Gaussian filtering, decouple the feature sequence of the rider's rational level into two parts: the main evolution pattern and the noise;

[0035] S303: For the decoupled main evolution pattern, perform time series modeling using polynomial regression; for the decoupled noise part, construct a Gaussian process based on the utility entropy for modeling;

[0036] S304: Estimate the current rational level noise using the utility entropy in the real-time scene information, and combine it with the time series prediction result to jointly generate the expected estimate of the rational level.

[0037] Furthermore, the calculation process of the utility entropy includes:

[0038] Sort the set of expected candidate trajectories based on the utility values of the candidate trajectories;

[0039] Use the equal-width method to discretize the utility values and divide them into multiple utility intervals;

[0040] Calculate the probability of each utility interval according to the number of candidates;

[0041] Calculate the utility entropy of the set of expected candidate trajectories, and the corresponding calculation expression is:

[0042]

[0043] In the formula, E d is the utility entropy, and P(i) is the probability of the utility interval i.

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] (1) The present invention proposes a non-motor vehicle adaptive trajectory prediction framework based on online update of rider decision-making characteristics. This framework can online identify and update the parameters of the rider's radical level and rational level, effectively overcoming the limitations of existing static pre-training models in dealing with heterogeneous riders, thereby significantly improving the prediction ability of the model for heterogeneous riders in dynamic applications.

[0046] (2) The present invention proposes a method for online recognition and dynamic update for the aggressive level and rational level of cyclists. Traditional models often have difficulty accurately predicting the trajectories of highly heterogeneous dynamic cyclists, especially cyclists with different types and behavior preferences. The present invention identifies their personalized characteristics through a prediction feedback mechanism, improving the model's ability to explain the decision-making motivation of cyclists in trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic flowchart of an adaptive trajectory prediction method for non-motor vehicles provided in an embodiment of the present invention;

[0048] Figure 2 is a diagram of a method for identifying the aggressive level provided in an embodiment of the present invention;

[0049] Figure 3 is a diagram of the prediction trajectory and the recognition result of behavioral characteristics provided in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and illustrated herein generally can be arranged and designed in a variety of different configurations.

[0051] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0052] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0053] As Figure 1 shown, the present invention provides an adaptive trajectory prediction method for non-motor vehicles, including the following steps:

[0054] Baseline prediction module S1: Based on real-time scene conditions, generate a set of expected candidate trajectory sets for the cyclist; evaluate and rank each trajectory through a utility function; finally select one of the trajectories as the prediction output at the current moment.

[0055] Feature Online Recognition Module S2: By analyzing the deviation between the actual observed value and the predicted value, identify and determine the personalized features of the rider, including the rider's aggression level and rationality level, to ensure that the framework can accurately capture their behavioral differences.

[0056] Parameter Online Update Module S3: Based on the known feature sequence, perform pattern modeling to explore the law of feature evolution. By deeply learning the feature change trend, dynamically adjust the model parameters so that the prediction model can more comprehensively adapt to the personalized features of the rider and provide optimized parameter support for the prediction of the next moment.

[0057] Personalized Prediction and Prediction Feedback Module S4: Combine with the prediction module to achieve personalized prediction for individual riders, and feedback the prediction results to the feature online recognition module to form a closed-loop optimization mechanism;

[0058] The implementation process of the prediction module includes three key steps: trajectory generation, utility evaluation, and sorting and selection.

[0059] As a preferred implementation, in the trajectory generation stage, the model generates a set of kinematically feasible candidate trajectory sets according to the real-time scene information. The specific steps are as follows: First, based on the 85th percentile of the maximum acceleration of non-motor vehicles, determine a set of feasible trajectory end regions and discretize them into multiple endpoints; then, use the cubic spline interpolation method to generate a complete set of candidate trajectories covering these endpoints.

[0060] The real-time scene information includes the non-motor vehicle to be measured, the positions, speeds, and accelerations of other traffic participants, as well as environmental information.

[0061] The kinematically feasible end region is calculated as follows. The boundaries of this region in the motion direction (X-axis) and the lateral direction (Y-axis) are respectively limited by the farthest and nearest reachable positions. Based on the basic assumption of constant acceleration, the above boundaries can be calculated by the following formula.

[0062] Ψ={x,y},x∈(x min ,x max )∩y∈(y min ,y max )

[0063]

[0064] In the formula, Ψ represents the reachable end region. x and y represent the coordinates of the end points in the region. x max ,x min ,y max and y min are the farthest and nearest possible positions of the bicycle along the lateral and longitudinal directions; x 0 、y0 , and are to study the current position and speed of the non-motor vehicle along the moving direction and laterally; and are the maximum positive and negative acceleration values along the moving direction and laterally.

[0065] The discrete sampling is specifically to perform spatial discrete sampling based on the diamond occupancy.

[0066] The feature online recognition module includes the online recognition of the rider's aggressive level and rational level.

[0067] As a preferred implementation manner, the rider's aggressive level is recognized by the following method:

[0068] First, a reasonable utility function is introduced to describe the heterogeneous trade-off between efficiency and safety for the rider. The expression of this utility function is as follows:

[0069]

[0070] In the formula, U(S i ) represents the utility value of candidate trajectory i, S i represents the coordinate matrix of trajectory i, and i ∈ {1, 2,... n}. n represents the total number of candidate trajectories. and respectively represent the risk value and efficiency value of the candidate trajectory. Assume that the trajectory has m trajectory points. The trajectory risk is quantified as the product of the average risk value E j and the path length d path . The risk field theory is used to describe the risk of each trajectory point, including the potential energy field and the dynamic field (i.e., E r and E v ). The trajectory efficiency is quantified as the mapped length of the trajectory in the forward direction. is the risk - efficiency preference, representing the rider's aggressive level, and when is close to 1, the rider tends to pursue efficiency (high aggressive level), and vice versa, tends to avoid risks (low aggressive level).

[0071] The online recognition method of the aggressive level uses the theoretically optimal trajectory as the representation of the true trajectory. The rider's theoretically optimal trajectory refers to the candidate trajectory in the candidate trajectory set that is most similar to the observed trajectory (i.e., has the lowest average distance error), as Figure 2 shown, and the expression is:

[0072]

[0073] In the formula, The radical level parameter S representing the theoretically optimal trajectory in the candidate trajectory set opt is the corresponding optimal trajectory. and are the coordinate matrices of the predicted trajectory and the observed trajectory, respectively. and dB represent the coordinates of the predicted trajectory point and the observed trajectory point at time step j. D(·) represents the calculation method of measuring trajectory similarity, defined as the average distance error. When the rider selects based on the candidate trajectory set, a utility comparison method is adopted. The value characterizes the weight of the safety term in the rider's utility, and its value ranges from 0 to 1. At the same time, the weight of the efficiency term is then can represent the radical level of the rider.

[0074] Here, the safety term in the utility function is characterized by "risk avoidance", so the utility U r has a superscript of "minus sign".

[0075] As a preferred implementation manner, the online update of the radical level is achieved through the following steps:

[0076] S201: Update the rider's radical level sequence based on the observed trajectory and the predicted feedback trajectory;

[0077] S202: Construct an evolution model through pattern learning based on the existing radical level sequence;

[0078] S203: Estimate the radical level parameter at a future moment based on the evolution model and apply it to the prediction model to enter the next iteration step.

[0079] As a preferred implementation manner, the evolution model describes the evolution process of the rider's radical level sequence, including two parts: the evolution pattern and the noise. The model decouples the sequence through Gaussian filtering, thereby separating the evolution pattern and the noise, and modeling them separately. Its expression form is:

[0080]

[0081] In the formula, represents the radical level parameter at time t, including the evolution main pattern Φ(t) and the noise sum. Φ(t) is a function with temporal regularity and can be fitted based on time through polynomial regression. The of the radical level parameter obeys a Gaussian distribution with a mean of 0 and a variance of

[0082] Further, the rational level of a cyclist is defined as the ratio of the individual utility selected by the cyclist in decision-making to the theoretical optimal individual utility, which is characterized by the following function:

[0083]

[0084] In the formula, λ is the characteristic parameter of the rational level of the cyclist in a certain decision-making, and utility evaluation is used for decision-making. U opt and U real respectively represent the utility values of the theoretical optimal candidate individual and the actually selected individual. The significance of the utility ratio lies in characterizing the possibility that the cyclist selects the theoretical optimal individual in decision-making.

[0085] As a preferred implementation, the rational level is defined as the ratio of the selected ranking to the difference of candidate trajectories, and online identification is realized through the following method, and the expression is:

[0086]

[0087] In the formula, λ t represents the characteristic parameter of the rational level. N target represents the ranking of the selected trajectory in the candidate trajectory evaluation. N all represents the number of candidates. When λ t is close to 1, it tends to be a selection based on the principle of maximizing utility, which means that they are cyclists with a completely rational level. On the contrary, it represents a bounded rational decision-making, and the degree of irrationality is related to the difficulty of ranking the utility of the candidate set.

[0088] The online update of the cyclist's rational level is based on a rational level learning and estimation method that integrates historical behavior characteristics and online scenario characteristics, including the following main steps:

[0089] S301: Continuously collect the utility entropy sequence and the rational level characteristic sequence as the basic data for modeling.

[0090] S302: Through Gaussian filtering processing, decouple the rational level characteristic sequence of the cyclist into two parts: the main evolution pattern and the noise.

[0091] S303: For the decoupled evolution pattern, use polynomial regression for time series modeling; for the noise part, build a Gaussian process based on utility entropy for modeling.

[0092] S304: Use the utility entropy in the online scenario to estimate the current rational level noise, and combine it with the time series prediction result to jointly generate an expected estimate.

[0093] As a preferred implementation, the expression of the above modeling method is:

[0094] λ(t) = Ω(t) + ελ [E d (t)]

[0095]

[0096] In the formula, λ(t) represents the parameter of the rational level at time step t, which is expressed as the sum of the main evolution mode Ω(t) and the noise ε λ [E d (t)]. Ω(t) is the main evolution mode of the rational level changing with time, and is fitted by polynomial regression. The rational level noise ε λ [E d (t) obeys the Gaussian distribution based on utility entropy, with a mean of 0 and a variance of

[0097] Furthermore, as a preferred implementation manner, the difficulty of the rider's decision-making scenario is characterized by utility entropy. When there are large differences in utility between candidate trajectories, the trajectory sorting is relatively easy, showing a relatively ordered state, and its entropy value is small; on the contrary, when the utility differences between candidate trajectories are small, the trajectory sorting becomes difficult, showing a relatively disordered state, and its entropy value is large.

[0098] As a preferred implementation manner, the utility entropy is calculated through the following steps:

[0099] S401 Utility sorting: Sort the candidate utility set based on the numerical size of the utility;

[0100] S402 Equal-width discretization: Discretize the utility values using the equal-width method and divide them into multiple utility intervals;

[0101] S403 Probability calculation: Calculate the probability of each utility interval according to the candidate numbers;

[0102] S404 Utility entropy calculation: Through the entropy value calculation formula of the utility distribution:

[0103]

[0104] In the formula, E d represents the entropy value of the utility distribution; P(i) represents the probability of the utility interval i.

[0105] Combining the above preferred implementation manners, multiple relatively optimal implementation manners can be obtained. The specific implementation process of an optimal implementation manner is described below.

[0106] Example 1

[0107] In this embodiment, the verification data was collected at the intersection of Xianxia Road and Jianhe Road in Shanghai. The length of the intersection in the east-west direction is 50 m, and the upstream and downstream sections are separated from the traffic flow by marking isolation machines. The width of the non-motor vehicle lane is 3.5 m. The video acquisition device is set on a high-rise building next to the intersection to obtain an aerial view. The video acquisition time is the evening rush hour (16:30 PM - 17:30 PM). After obtaining the video data, this paper used video trajectory extraction software to extract the trajectories of non-motor vehicles and other traffic participants, and the extraction step size was 0.12 s. Finally, 680 non-motor vehicle trajectories were obtained, and the labels included trajectory coordinates, speed, acceleration, curvature, and vehicle type. In the embodiment, the prediction time is 3 s (25 steps).

[0108] This embodiment provides a non-motor vehicle adaptive trajectory prediction method, including the following steps:

[0109] Baseline prediction module S1: Based on real-time scene conditions, generate a set of expected candidate trajectory sets for cyclists; evaluate and sort each trajectory through a utility function; finally, select one of the trajectories as the prediction output at the current moment;

[0110] Feature online recognition module S2: By analyzing the deviation between the actual observation value and the predicted value, identify and determine the personalized features of the cyclist, including the aggressive level and rational level of the cyclist, to ensure that the framework can accurately capture their behavioral differences;

[0111] Parameter online update module S3: Based on the known feature sequence, perform pattern modeling to explore the law of feature evolution. By deeply learning the feature change trend, dynamically adjust the model parameters, so that the prediction model can more comprehensively adapt to the personalized features of the cyclist and provide optimized parameter support for the prediction of the next moment;

[0112] Personalized prediction and prediction feedback module S4: Combine the prediction module to achieve personalized prediction for individual cyclists, and feedback the prediction results to the feature online recognition module to form a closed-loop optimization mechanism;

[0113] The implementation process of the prediction module includes three key steps: trajectory generation, utility evaluation, and sorting and selection.

[0114] As a preferred implementation method, in the trajectory generation stage, the model generates a set of kinematically feasible candidate trajectory sets according to real-time scene information. The specific steps are as follows: First, based on the 85th percentile of the maximum acceleration of non-motor vehicles, determine a set of feasible trajectory end regions and discretize them into multiple endpoints; then, use the cubic spline interpolation method to generate a complete set of candidate trajectories covering these endpoints.

[0115] The real-time scene information includes the non-motor vehicle to be measured, the positions, speeds, and accelerations of other traffic participants, as well as environmental information.

[0116] The kinematically feasible end region is calculated as follows. The boundaries of this region in the motion direction (X-axis) and the lateral direction (Y-axis) are respectively limited by the farthest and nearest reachable positions. Based on the basic assumption of constant acceleration, the above boundaries can be calculated by the following formulas.

[0117] Ψ = {x,y}, x ∈ (x min , x max ) ∩ y ∈ (y min , y max )

[0118]

[0119] In the formula, Ψ represents the reachable end region. x and y represent the coordinates of the end points within the region. x max , x min , y max and y min are the possible farthest and nearest positions of the bicycle along the lateral and longitudinal directions; x 0 , y 0 , and are the current positions and speeds of the non-motor vehicle along the motion direction and the lateral direction; and are the maximum positive and negative acceleration values along the motion direction and the lateral direction.

[0120] The discrete sampling is specifically to perform spatial discretization sampling based on the diamond occupancy.

[0121] The feature online recognition module includes the online recognition of the rider's aggressive level and rational level.

[0122] The rider's aggressive level is recognized by the following method:

[0123] First, a reasonable utility function is introduced to describe the heterogeneous trade-off between efficiency and safety for the rider. The expression of this utility function is as follows:

[0124]

[0125] In the formula, U(S i ) represents the utility value of the candidate trajectory i, S i represents the coordinate matrix of the trajectory i, i ∈ {1,2,...n}. n represents the total number of candidate trajectories. and They represent the risk value and efficiency value of the candidate trajectory respectively. Suppose the trajectory has m trajectory points. The trajectory risk is quantified as the average risk value E of each trajectory point j multiplied by the path length d path . The risk field theory is used to describe the risk of each trajectory point, including the potential energy field and the dynamic field (i.e., E r and E v ). The trajectory efficiency is quantified as the mapped length of the trajectory in the forward direction. is the risk - efficiency preference, representing the aggressiveness level of the rider, and when is close to 1, the rider tends to pursue efficiency (high aggressiveness level), otherwise tends to avoid risks (low aggressiveness level). For the risk field, the parameter settings of the risk quantification values are as follows: φ1 = 0.1, μ1 = 2.2, φ2 = 5, μ2 = 0.4, and R j = 1. The parameter settings of the existing risk field are k1 = 1, k2 = 0.05, and G = 0.1.

[0126] The online identification method of the aggressiveness level uses the theoretical optimal trajectory as the representation of the true trajectory. The theoretical optimal trajectory of the rider refers to the candidate trajectory in the candidate trajectory set that is most similar to the observed trajectory (i.e., has the lowest average distance error), and the expression is:

[0127]

[0128] In the formula, represents the aggressiveness level parameter of the theoretical optimal trajectory in the candidate trajectory set, and S opt is the corresponding optimal trajectory. and are the coordinate matrices of the predicted trajectory and the observed trajectory respectively. and represent the coordinates of the predicted trajectory point and the observed trajectory point in the time step j respectively. D(·) represents the calculation method for measuring the trajectory similarity, which is defined as the average distance error.

[0129] The online update of the aggressiveness level is achieved through the following steps:

[0130] S201: Update the aggressiveness level sequence of the rider based on the observed trajectory and the predicted feedback trajectory;

[0131] S202: Construct an evolution model through pattern learning based on the existing aggressiveness level sequence;

[0132] S203: Estimate the aggressiveness level parameter at the future moment based on the evolution model and apply it to the prediction model, and enter the next iteration step.

[0133] The evolutionary model describes the evolution process of the sequence of cyclists' aggressiveness levels, including two parts: the evolutionary pattern and the noise. This model decouples the sequence through Gaussian filtering to separate the evolutionary pattern and the noise, and models them separately. Its expression form is:

[0134]

[0135] In the formula, represents the aggressiveness level parameter at time t, including the evolutionary main pattern Φ(t) and the noise The sum. Φ(t) is a function with temporal regularity and can be fitted based on time through polynomial regression. The of the aggressiveness level parameter obeys a Gaussian distribution with a mean of 0 and a variance of . In the embodiment, the Gaussian filtering step size and the polynomial exponent are taken as 5 and 3 respectively. In the embodiment, the initial value of the aggressiveness level is set to 0.5.

[0136] The rational level of a cyclist is defined as the ratio of the individual utility selected by the cyclist in the decision-making to the theoretical optimal individual utility, and is characterized by the following function:

[0137]

[0138] In the formula, λ is the rational level characteristic parameter of the cyclist in a certain decision-making, and utility evaluation is used for decision-making. U opt and U real represent the utility values of the theoretical optimal candidate individual and the actually selected individual respectively. The significance of the utility ratio lies in characterizing the possibility that the cyclist selects the theoretical optimal individual in the decision-making.

[0139] The rational level is defined as the ratio of the selected ranking to the difference of the candidate trajectories, and is realized for online identification through the following method. The expression is:

[0140]

[0141] In the formula, λ t represents the rational level characteristic parameter. N target represents the ranking of the selected trajectory in the candidate trajectory evaluation. N all represents the number of candidates. When λ t is close to 1, it tends to be a selection based on the principle of maximizing utility, which means they are cyclists with a completely rational level. On the contrary, it represents a limited rational decision-making, and the degree of irrationality is related to the difficulty of ranking the utility of the candidate set. In the embodiment, the initial value of the aggressiveness level is set to 1.

[0142] The online update of the cyclist's rational level is based on a rational level learning and estimation method that integrates historical behavior characteristics and online scenario characteristics, including the following main steps:

[0143] S301: Continuously collect the utility entropy sequence and the rational level feature sequence as the basic data for modeling.

[0144] S302: Through Gaussian filtering, decouple the rational level feature sequence of the rider into two parts: the main evolutionary pattern and the noise.

[0145] S303: For the decoupled evolutionary pattern, perform time series modeling using polynomial regression; for the noise part, construct a Gaussian process based on utility entropy for modeling.

[0146] S304: Use the utility entropy in the online scenario to estimate the current rational level noise, and combine it with the time series prediction result to jointly generate the expected estimate.

[0147] The expression of the above modeling method is:

[0148] λ(t) = Ω(t) + ε λ [E d (t)]

[0149]

[0150] In the formula, λ(t) represents the parameter of the rational level at time step t, which is expressed as the sum of the main evolutionary pattern Ω(t) and the noise ε λ [E d (t)]. Ω(t) is the main evolutionary pattern of the rational level changing with time, and is fitted by polynomial regression. The rational level noise ε λ [E d (t)] follows a Gaussian distribution based on utility entropy, with a mean of 0 and a variance of

[0151] The difficulty of the rider's decision-making scenario is characterized by utility entropy. When there are large differences in utility between candidate trajectories, the trajectory ranking is relatively easy, showing a relatively ordered state with a small entropy value; on the contrary, when the utility differences between candidate trajectories are small, the trajectory ranking becomes difficult, showing a relatively disordered state with a large entropy value. The utility entropy is calculated through the following steps:

[0152] S401 Utility ranking: Rank the candidate utility set based on the numerical size of the utility;

[0153] S402 Equal-width discretization: Discretize the utility values using the equal-width method and divide them into multiple utility intervals;

[0154] S403 Probability calculation: Calculate the probability of each utility interval according to the candidate numbers;

[0155] S404 Utility Entropy Calculation: Through the entropy value calculation formula of the utility distribution:

[0156]

[0157] In the formula, E d represents the entropy value of the utility distribution; P(i) represents the probability of the utility interval i.

[0158] Based on the above steps, this embodiment realizes the adaptive prediction of non-motor vehicles. The specific results are as follows: The test results of the adaptive model are shown in Table 1. Compared with the baseline model, the prediction error of the adaptive model is smaller, indicating that it has higher reliability in predicting the movement of non-motor vehicles. In addition, the prediction error of the model that only updates the aggressive level is lower than that of the model that only updates the rational level, which shows that the change in the aggressive level has a more significant impact on the movement of non-motor vehicles.

[0159] To further analyze the advantages of the online update of the rider's decision variables, the efficiency and error improvement rate on samples of different complexities are calculated. Tables 2 and 3 show the efficiency and improvement performance when the aggressive level, rational level, and their combined application are considered. The results show that as the complexity of the prediction scenario increases, the online update effects of both rider decision-making characteristics are enhanced. In the regular scenario (all samples and samples with an ADE threshold greater than 0.4 meters), the aggressive level shows higher effectiveness, while the rational level is superior in terms of the error reduction rate. In more complex scenarios (samples with an ADE threshold greater than 0.8 meters and 1.2 meters), the aggressive level is superior to the rational level in both the efficiency and error improvement rate indicators.

[0160] In addition, in the continuous prediction of individuals, the prediction effect is as Figure 3 shown.

[0161] Table 1: Comparison of Prediction Errors

[0162] Model ADE / m FDE / m Baseline model 0.50 1.38 Aggressive level learning model 0.38 1.05 Rational level learning model 0.39 1.08 Comprehensive learning model 0.37 1.03

[0163] Table 2: Effectiveness of the Adaptive Prediction Model in Different Complex Scenarios

[0164]

[0165] —The error threshold is based on the baseline model

[0166] Table 3: Error Improvement Rate of the Adaptive Prediction Model in Different Complex Scenarios

[0167]

[0168] —The error threshold is based on the baseline model

[0169] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art shall fall within the protection scope determined by the claims.

Claims

1. An adaptive trajectory prediction method for non-motor vehicles, characterized in that, It includes the following steps: S1: Based on real-time scenario information, generate a set of expected candidate trajectories for the rider, and evaluate each trajectory in the set of expected candidate trajectories through a preset utility function, and select the theoretically optimal trajectory as the predicted value of the trajectory at the current moment; S2: Obtain the actual observation value at the current moment, calculate the deviation from the corresponding predicted value, identify and determine the personalized characteristics of the rider, and the personalized characteristics include the aggressive level and rational level of the rider; S3: Perform pattern modeling according to the determined sequence of personalized characteristics to obtain the feature evolution law, so as to adjust the parameters of the pre-constructed trajectory prediction model; S4: Use the adjusted trajectory prediction model to perform personalized prediction on the rider individual, and return to step S2 according to the obtained predicted value to update the personalized characteristics of the rider, forming a closed-loop optimization.

2. The adaptive trajectory prediction method for non-motor vehicles according to claim 1, wherein The generation process of the set of expected candidate trajectories in step S1 includes the following steps: S101: Based on real-time scenario information, determine the feasible end area of the rider according to the motion law; S102: Discretely sample the feasible end area into a set of endpoints; S103: According to the obtained endpoints, use the trajectory interpolation method to generate corresponding candidate trajectories respectively, and generate a set of kinematically feasible expected candidate trajectory sets.

3. The adaptive trajectory prediction method for non-motor vehicles according to claim 2, characterized in that, The real-time scenario information includes the position, speed and acceleration information of the rider to be measured and other traffic participants, as well as environmental information.

4. An adaptive trajectory prediction method for non-motor vehicles according to claim 2, characterized in that, The feasible end area is determined by calculating the farthest reachable position and the nearest reachable position in the motion direction and the lateral direction.

5. The adaptive trajectory prediction method for non-motor vehicles according to claim 1, wherein The expression of the utility function is: where U(S i ) is the utility value of candidate trajectory i, S i is the coordinate matrix of trajectory i, n is the total number of candidate trajectories, and represent the risk value and efficiency value of the candidate trajectory respectively, is the risk - efficiency preference, representing the aggressive level of the rider, 6. The non-motor vehicle self-adaptive trajectory prediction method according to claim 1, wherein The calculation expression of the identification process of the aggressive level is: In the formula, is the aggressiveness level parameter of the theoretically optimal trajectory in the expected candidate trajectory set, S opt is the corresponding optimal trajectory, and respectively represent the risk value and efficiency value of the trajectory S opt , is the weight of the safety term in the rider's utility, and are the coordinate matrices of the trajectory prediction value and the trajectory observation value respectively, n is the total number of candidate trajectories, and respectively represent the coordinates of the predicted trajectory point and the observed trajectory point at time step j, and D(·) is the calculation method for measuring the similarity of the trajectories.

7. An adaptive trajectory prediction method for non-motor vehicles according to claim 6, characterized in that, The online update process of the aggressive level includes: S201: Based on the actual observation value of the trajectory and the corresponding predicted value, update the sequence of the aggressive level of the rider; S202: Based on the existing sequence data of the aggressive level, perform pattern learning to construct an aggressive level evolution model; S203: Based on the constructed aggressive level evolution model, estimate the aggressive level parameters at future moments and use them to adjust the parameters of the trajectory prediction model.

8. A non-motor vehicle adaptive trajectory prediction method according to claim 1, wherein The calculation expression of the identification process of the rational level is: where λ t represents the rational level characteristic parameter, N target represents the ranking of the selected trajectory in the candidate trajectory evaluation, and N all represents the number of candidates.

9. The adaptive trajectory prediction method for non-motor vehicles according to claim 1, wherein The rational level is defined as the ratio of the individual utility selected by the rider in the decision-making to the theoretically optimal individual utility; The learning and estimation process of the rational level includes: S301: Continuously collect the sequence of utility entropy and the sequence of characteristics of the rational level; S302: Through Gaussian filtering processing, decouple the sequence of characteristics of the rational level of the rider into two parts: the main evolution pattern and the noise; S303: For the decoupled main evolution pattern, perform time series modeling using polynomial regression; for the decoupled noise part, construct a Gaussian process based on utility entropy for modeling; S304: Use the utility entropy in the real-time scenario information to estimate the current rational level noise, and combine it with the time series prediction result to jointly generate the expected estimate of the rational level.

10. A non-motor vehicle adaptive trajectory prediction method according to claim 9, characterized in that, The calculation process of the utility entropy includes: Sort the set of expected candidate trajectories based on the magnitude of the utility value of the candidate trajectories; Use the equal-width method to discretize the utility values and divide them into multiple utility intervals; Calculate the probability of each utility interval based on the candidate numbers; Calculate the utility entropy of the expected candidate trajectory set, and the corresponding calculation expression is: where E d is the utility entropy, and P(k) is the probability of the utility interval k.

Images