Vehicle behavior prediction method and device with task-free lifelong learning capability, equipment, medium and product

Abstract

The invention discloses a vehicle behavior prediction method and device with task-free lifelong learning ability, equipment, a medium and a product, and relates to the field of vehicle automatic driving, and the method comprises the steps: carrying out the real-time monitoring of a traffic environment, and obtaining dynamic and static traffic environment data; constructing a rapid memory module and a difference memory module to store dynamic and static traffic environment data; when a saturation state is reached, updating data by adopting a reservoir sampling method and a gradient selection sampling method respectively; and inputting data in the two modules into a vehicle behavior prediction model, training through a task-free memory playback loss function, and outputting a prediction trajectory of a target vehicle in a prediction time period. The vehicle behavior prediction model has a task-free lifelong learning ability, can continuously adapt to continuous changes of a traffic environment without depending on task switching information and additional retraining, always keeps high prediction precision, and effectively alleviates the problem of disastrous forgetting possibly occurring when a traditional model faces a new scene.

Description

Technical Field

[0001] The present application relates to the field of vehicle autonomous driving, and in particular to a vehicle behavior prediction method, device, equipment, medium and product with task-free lifelong learning capability. Background Art

[0002] With the rapid advancement of science and technology, intelligent transportation systems have become an important direction for future transportation development. In intelligent transportation systems, vehicle behavior prediction is one of the core technologies in multiple fields such as autonomous driving, intelligent traffic management and optimization. Through comprehensive analysis of various factors in the traffic environment, the future driving trajectory and behavior of the vehicle can be predicted, thereby providing decision support for autonomous driving vehicles and providing data reference for traffic management departments to achieve optimization of traffic flow and improvement of safety.

[0003] With the acceleration of urbanization and the continuous increase in the number of traffic participants, the modern traffic environment has become increasingly complex and changeable, which has put forward higher requirements for vehicle behavior prediction technology. In this environment, autonomous vehicles need to accurately predict the future behavior of surrounding vehicles in order to make reasonable behavior decisions and motion planning to ensure driving safety and efficiency.

[0004] In order to improve the prediction accuracy and model performance, the behavior prediction methods based on deep learning in the related art often choose to increase and enrich the number and coverage of scenes in the training data set. However, this type of method not only requires the model to have high complexity and long training convergence time, but also it is difficult to effectively use the knowledge and experience obtained from historical scenes and data training for efficient scene understanding and knowledge representation in continuously changing scenes. Therefore, when the model is trained on data of new scenes, its performance often drops sharply in the learned scenes, which is called "catastrophic forgetting" in the field of deep learning. Although a small number of methods use lifelong learning strategies to deal with this problem, such methods in the related art all rely on scene switching information to implement lifelong learning strategies. However, in the actual application of autonomous driving, different scenes are often difficult to identify and define in real time by the autonomous driving system, so scene switching information is difficult to obtain. In view of the limitations of the above-mentioned vehicle behavior prediction methods, the present application proposes a vehicle behavior prediction method with task-free lifelong learning capabilities. Summary of the invention

[0005] The purpose of this application is to provide a vehicle behavior prediction method, device, equipment, medium and product with task-free lifelong learning capability, which can realize the knowledge and experience accumulation of the vehicle behavior prediction model without relying on scene switching information and avoid catastrophic forgetting.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] In a first aspect, the present application provides a vehicle behavior prediction method with task-free lifelong learning capability, comprising:

[0008] The traffic environment is monitored to obtain dynamic and static traffic environment data of the current time period; the current time period includes the current moment and a first preset number of moments before the current moment; the dynamic and static traffic environment data includes: position information and speed information of all traffic participants in the current traffic environment, and road feature information; traffic participants include a target vehicle and vehicles within a preset range of the target vehicle;

[0009] Construct fast memory modules and difference memory modules with preset number threshold sizes;

[0010] If the fast memory module has not reached the first saturation state, the dynamic and static traffic environment data of the current time period are stored in the fast memory module; if the fast memory module has reached the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method; wherein the first saturation state is when the amount of data stored in the fast memory module reaches a preset quantity threshold;

[0011] If the difference memory module has not reached the second saturation state, the dynamic and static traffic environment data of the current time period is stored in the difference memory module; if the difference memory module reaches the second saturation state, the dynamic and static traffic environment data in the difference memory module is updated using a gradient selection sampling method; wherein the second saturation state is when the amount of data stored in the difference memory module reaches a preset quantity threshold;

[0012] The dynamic and static traffic environment data of the current time period in the fast memory module and the difference memory module are input into the vehicle behavior prediction model, the vehicle behavior prediction model is trained by the task-free memory playback loss function, and the predicted trajectory of the target vehicle in the prediction time period is output; the prediction time period includes a second preset number of moments after the current moment.

[0013] Optionally, if the fast memory module reaches the first saturation state, a reservoir sampling method is used to update the dynamic and static traffic environment data in the fast memory module, specifically including:

[0014] Generate a random number r1 between the interval [0, j-1], and compare the generated random number r1 with a preset number threshold;

[0015] If the generated random number r1 is less than the preset number threshold, the r1th dynamic and static traffic environment data stored in the fast memory module is replaced with the dynamic and static traffic environment data of the current time period;

[0016] If the generated random number r1 is greater than or equal to the preset quantity threshold, the fast memory module is not updated.

[0017] Optionally, if the difference memory module reaches the second saturation state, a gradient selection sampling method is used to update the dynamic and static traffic environment data in the difference memory module, specifically including:

[0018] Calculate the similarity score between the jth dynamic and static traffic environment data in the difference memory module and the preset data set; where, when j=1, the similarity score is set to 0.1; the preset data set is a preset number of dynamic and static traffic environment data randomly selected from the difference memory module except the jth dynamic and static traffic environment data; when j>1, the calculation formula of the similarity score is:

[0019]

[0020] Among them, q j represents the similarity score between the jth dynamic and static traffic environment data and the preset data set, g j represents the gradient of the loss function of the jth dynamic and static traffic environment data, g k Represents the loss function gradient of the kth dynamic and static traffic environment data in the preset data set, k = 1, 2, 3, ..., n; k ≠ ​​j, n represents the preset selection number;

[0021] The interval [0, 1] is The probability size is divided into a preset number threshold molecular interval; among which, P(j) represents the length of the j-th subinterval, M represents the preset number threshold, m=0,1,2,...,M-1;

[0022] Generate a random number r2 between 0 and 1, and when r2 falls within the j-th subinterval, use the j-th dynamic and static traffic environment data stored in the difference memory module as the dynamic and static traffic environment data to be replaced;

[0023] Generate a random number r3 between 0 and 1 and compare the random number r3 with q i / (q i +q j ) size; if r3 is smaller than q i / (q i +q j ), replace the dynamic and static traffic environment data to be replaced with the dynamic and static traffic environment data of the current time period; if r3 is greater than or equal to q i / (q i +q j ), the difference memory module is not updated; where q iIndicates the similarity score between the dynamic and static traffic environment data of the current time period and the preset data set.

[0024] Optionally, the calculation formula of the task-free memory replay loss function is:

[0025]

[0026] in, represents the task-free memory replay loss function, represents the training loss function, and They represent the fast memory module and the difference memory module respectively. and They represent the losses of the fast memory module and the difference memory module respectively, α and β are the weight coefficients of the first loss term and the second loss term respectively, and f θ represents the vehicle behavior prediction model, X is the dynamic and static traffic environment data input into the vehicle behavior prediction model, Y represents the actual trajectory of the target vehicle within the prediction time period, and X R ' and X D 'represents the dynamic and static traffic environment data stored in the fast memory module and the difference memory module, respectively, Y R ' indicates X R 'The actual trajectory of the target vehicle during the corresponding prediction period, Y D ' indicates X D 'The actual trajectory of the target vehicle during the corresponding prediction time period, Represents X R 'The predicted trajectory of the target vehicle within the corresponding prediction time period, Represents X R 'The predicted trajectory of the target vehicle within the corresponding prediction time period, represents the mathematical expectation, represents the mathematical expectation of the dynamic and static traffic environment data stored in the fast memory module and the corresponding actual trajectory of the target vehicle, represents the mathematical expectation of the dynamic and static traffic environment data stored in the difference memory module and the corresponding actual trajectory of the target vehicle, represents the mathematical expectation of the dynamic and static traffic environment data stored in the fast memory module and its corresponding target vehicle prediction trajectory, Represents the mathematical expectation of the dynamic and static traffic environment data stored in the difference memory module and its corresponding predicted trajectory of the target vehicle.

[0027] Optionally, the vehicle behavior prediction model specifically includes: an encoder module, a predictor module, and a regularization network module;

[0028] The encoder module is connected to the predictor module and the regularization network module respectively;

[0029] The encoder module is used to divide the dynamic and static traffic environment data into grid coordinate information, lane line information, target vehicle motion trajectory information, and motion trajectory information of vehicles within a preset range of the target vehicle;

[0030] The predictor module is used to obtain a first predicted trajectory of the target vehicle within a predicted time period based on grid coordinate information, lane line information, target vehicle motion trajectory information, and motion trajectory information of vehicles within a preset range of the target vehicle;

[0031] The regularized network module is used to obtain a second predicted trajectory of the target vehicle within a predicted time period according to the grid coordinate information, the lane line information and the target vehicle motion trajectory information;

[0032] The first predicted trajectory and the second predicted trajectory are merged to obtain a final predicted trajectory of the target vehicle within the prediction time period.

[0033] Optionally, the training loss function is calculated as:

[0034]

[0035]

[0036] in, is the L1 norm, represents the first predicted trajectory, represents the second predicted trajectory, Indicates that Linearly renormalize from 0 to 1.

[0037] In a second aspect, the present application provides a vehicle behavior prediction device with task-free lifelong learning capability, comprising:

[0038] The environment data acquisition module is used to monitor the traffic environment and obtain dynamic and static traffic environment data of the current time period; the current time period includes the current time and the first preset number of time before the current time; the dynamic and static traffic environment data includes: the position information and speed information of all traffic participants in the current traffic environment, and road feature information; traffic participants include target vehicles and vehicles within a preset range of the target vehicle;

[0039] A building block of a memory module, used to build a fast memory module and a difference memory module of a preset number threshold size;

[0040] The first data module is used to store the dynamic and static traffic environment data of the current time period into the fast memory module if the fast memory module has not reached the first saturation state; if the fast memory module reaches the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method; wherein the first saturation state is when the amount of data stored in the fast memory module reaches a preset quantity threshold;

[0041] A second data module is used to store the dynamic and static traffic environment data of the current time period into the difference memory module if the difference memory module has not reached the second saturation state; if the difference memory module has reached the second saturation state, the dynamic and static traffic environment data in the difference memory module is updated by using a gradient selection sampling method; wherein the second saturation state is when the amount of data stored in the difference memory module reaches a preset quantity threshold;

[0042] A predicted trajectory acquisition module is used to input the dynamic and static traffic environment data of the current time period in the fast memory module and the difference memory module into the vehicle behavior prediction model, train the vehicle behavior prediction model through the task-free memory playback loss function, and output the predicted trajectory of the target vehicle within the predicted time period; the predicted time period includes a second preset number of moments after the current moment.

[0043] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-mentioned vehicle behavior prediction methods with task-free lifelong learning capability.

[0044] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned vehicle behavior prediction methods with task-free lifelong learning capability.

[0045] In a fifth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned vehicle behavior prediction methods with task-free lifelong learning capability.

[0046] According to the specific embodiments provided in this application, this application has the following technical effects:

[0047] The present application provides a vehicle behavior prediction method, device, equipment, medium and product with task-free lifelong learning capability. By real-time monitoring of the traffic environment, dynamic and static traffic environment data within the current time period are obtained, including the location information, speed information and road feature information of all traffic participants. This solves the problem that the traditional vehicle behavior prediction method only relies on limited historical data and cannot fully reflect the current traffic conditions, and realizes comprehensive perception and real-time update of the traffic environment. A fast memory module and a difference memory module with a preset number of threshold sizes are constructed, and the data is updated using the reservoir sampling and gradient selection sampling methods according to the saturation state of the module, which solves the storage pressure and low computing efficiency caused by the increasing amount of data in the long-term learning process of the model, and realizes effective management and efficient use of data. At the same time, the fast memory module captures recent and frequently changing traffic information, and the difference memory module records long-term and significantly different traffic scenes, which further improves the adaptability and prediction accuracy of the model. The data in the fast memory module and the difference memory module are input into the vehicle behavior prediction model, and the task-free memory replay loss function is used for training, which solves the catastrophic forgetting problem that may occur in traditional models in new scenarios and enables the model to maintain high prediction accuracy without relying on task switching and additional retraining. The task-free memory replay loss function can use the stored memory data to optimize the model parameters, so that the model can better adapt to the continuous changes in the traffic environment.

[0048] In summary, this application solves the problems of incomplete data, high storage pressure, low computational efficiency, and decreased prediction accuracy in new scenarios in existing vehicle behavior prediction methods by real-time monitoring of the traffic environment, building a dual memory module and adopting an effective data update method, and using a task-free memory playback loss function to train the prediction model. It achieves comprehensive perception of the traffic environment, effective management and efficient use of data, and the ability of the model to maintain high prediction accuracy without relying on additional training. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 A flowchart of a vehicle behavior prediction method with task-free lifelong learning capability provided in one embodiment of the present application;

[0051] Figure 2A schematic diagram of functional modules of a vehicle behavior prediction device with task-free lifelong learning capability provided in one embodiment of the present application;

[0052] Figure 3 A schematic diagram of the structure of a vehicle behavior prediction model provided in one embodiment of the present application;

[0053] Figure 4 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0054] First, some technical terms involved in the embodiments of the present application are introduced.

[0055] Catastrophic forgetting refers to the phenomenon that after the vehicle behavior prediction model is trained with new scene data, the prediction accuracy drops significantly in the test of the learned historical scenes due to differences in road structure, traffic rules, types and behaviors of traffic participants in different scenes.

[0056] Task,refers to the prediction of vehicle behavior in a certain traffic scenario, which is divided according to the scenario.

[0057] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0058] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0059] In an exemplary embodiment, Figure 1 As shown, a vehicle behavior prediction method with task-free lifelong learning capability is provided, which is executed by a computer device, specifically, it can be executed by a computer device such as a terminal or a server alone, or it can be executed by a terminal and a server together. In the embodiment of the present application, the following steps 101 to 105 are included. Among them:

[0060] Step 101, monitor the traffic environment and obtain dynamic and static traffic environment data of the current time period; the current time period includes the current moment and the first preset number of moments before the current moment; the dynamic and static traffic environment data include: position information and speed information of all traffic participants in the current traffic environment, and road feature information; traffic participants include target vehicles and vehicles within a preset range of the target vehicle.

[0061] In a specific embodiment of the present application, the dynamic and static traffic environment data including the position, speed information and road feature information of traffic participants in step 101 include: using sensors to collect traffic scene data. A coordinate origin O is selected in the acquired scene to establish an xOy plane rectangular coordinate system, and the positions of each traffic participant include the coordinate point (x, y) of the center position of all traffic participants including the vehicle in the xOy rectangular coordinate system, in meters, accurate to three decimal places; the speed information is contained in the xOy plane rectangular coordinate system, and the speed value of each traffic participant relative to the longitudinal and lateral directions of the vehicle, in meters / second; the road feature information refers to the lane line information and road boundary information identified in the picture collected by the sensor.

[0062] As an optional implementation, dynamic and static traffic environment data including the position, speed information and road feature information of traffic participants are obtained and data preprocessing is performed. The preprocessing specifically includes: extracting the horizontal and vertical position coordinates, horizontal and vertical speed information of each traffic participant (vehicle, pedestrian, cyclist), and processing them into vector samples with a sample time span of 4 seconds with the road centerline coordinates and boundary line information. The first 1 second of each sample is used as the observation input, and the prediction time window is the last 3 seconds, that is, the last 3 seconds is used as the actual trajectory Y of the target vehicle in supervised learning. Each sample contains the motion characteristics and road characteristics of a selected main vehicle and traffic participants within a certain range near it. The motion characteristics are the horizontal and vertical position and horizontal and vertical speed information under different time frames; the road characteristics are the corresponding lane line coordinates. All features are represented in vector form.

[0063] Step 102, constructing a fast memory module and a difference memory module of a preset number threshold size.

[0064] Step 103, if the fast memory module has not reached the first saturation state, the dynamic and static traffic environment data of the current time period is stored in the fast memory module; if the fast memory module reaches the first saturation state, the dynamic and static traffic environment data in the fast memory module is updated using a reservoir sampling method; wherein the first saturation state is that the amount of data stored in the fast memory module reaches a preset quantity threshold.

[0065] As an optional implementation, step 103 specifically includes:

[0066] If the fast memory module reaches the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method, specifically including:

[0067] Generate a random number r1 between the interval [0, j-1], and compare the generated random number r1 with a preset quantity threshold.

[0068] If the generated random number r1 is less than the preset quantity threshold, the r1-th dynamic and static traffic environment data stored in the fast memory module is replaced with the dynamic and static traffic environment data of the current time period.

[0069] If the generated random number r1 is greater than or equal to the preset quantity threshold, the fast memory module is not updated.

[0070] The above reservoir sampling is characterized in that the memory data samples stored in the fast memory module are random, and sampling is equally likely from a data stream of unknown total length.

[0071] Step 104, if the difference memory module does not reach the second saturation state, the dynamic and static traffic environment data of the current time period is stored in the difference memory module; if the difference memory module reaches the second saturation state, the gradient selection sampling method is used to update the dynamic and static traffic environment data in the difference memory module; wherein, the second saturation state is that the amount of data stored in the difference memory module reaches the preset quantity threshold.

[0072] As an optional implementation manner, step 104 specifically includes:

[0073] First, assign a score for measuring the similarity of data samples to each data sample stored in the difference memory module, and the difference memory module will select samples from the data stream for storage based on the principle of minimizing the similarity between the stored samples. For the first training sample in the data stream S, initialize its similarity score to 0.1 and store it in the difference memory module. Let the similarity score of the j-th sample be q j , randomly select n (n < M) samples from the remaining memory data samples, compare the gradient cosine similarity between the j-th data sample and the n samples in turn, select the maximum value of the n cosine similarities, and add 1 (to ensure the score is non-negative) as the score. Wherein, select an n less than the upper limit of the module capacity, and when the number of samples stored in the memory module is less than n, all the stored samples are taken for calculation.

[0074] The similarity score calculation is based on the sample gradient cosine similarity, specifically as follows: Let the gradient of the loss function corresponding to the j-th sample be g j , and the gradients of the loss functions corresponding to the randomly selected n data samples be g k (k = 1, 2, 3,... n; k ≠ j), then there is:

[0075]

[0076] Among them, q j represents the similarity score between the j-th dynamic and static traffic environment data and the preset data set, g jrepresents the gradient of the loss function of the jth dynamic and static traffic environment data, g k Represents the loss function gradient of the kth dynamic and static traffic environment data in the preset data set, k = 1, 2, 3, ..., n; k ≠ ​​j, n represents the preset selection number.

[0077] When the difference memory module is not saturated, that is, When the dynamic and static traffic environment data of the current time period is stored in the difference memory module, the similarity score q j When the difference memory module is saturated, it will consider whether to replace the stored samples with the dynamic and static traffic environment data of the current time period based on the similarity score. The specific replacement strategy is as follows: When j <1, with probability The jth dynamic and static traffic environment data stored in the difference memory module is selected as the candidate sample to be replaced. This step is implemented by the following process: First, the interval [0, 1] is replaced according to The probability size is divided into M sub-intervals, among which, The length of the jth subinterval is Generate a uniform random number r2 between the interval [0, 1], and when r2 falls within the jth interval, select the jth data sample as the candidate sample to be replaced.

[0078] On the basis of determining the candidate samples, a new random number r3 is further generated in the interval [0, 1], and the random number r3 is compared with q i / (q i +q j ), q i Indicates the similarity score between the dynamic and static traffic environment data of the current time period and the preset data set.

[0079] When i / (q i +q j ), the jth dynamic and static traffic environment data in the difference memory module is replaced with the dynamic and static traffic environment data of the current time period, otherwise, the dynamic and static traffic environment data of the current time period used for model training is discarded, and the difference memory module is not updated. The sampling method based on gradient selection can maximize the difference between data samples in the difference memory module and improve the data diversity of the difference memory module.

[0080] Step 105, input the dynamic and static traffic environment data of the current time period in the fast memory module and the difference memory module into the vehicle behavior prediction model, train the vehicle behavior prediction model through the task-free memory playback loss function, and output the predicted trajectory of the target vehicle within the prediction time period; the prediction time period includes a second preset number of moments after the current moment.

[0081] In another exemplary embodiment of the present application, the vehicle behavior prediction model is denoted as f θ , the dynamic and static traffic environment data of the input model is (X,Y), where X is the observation value of the dynamic and static traffic environment data, and Y is the actual trajectory of the target vehicle during the prediction time period. X can be expressed as:

[0082] X={Ο τ ,Ε τ |tt obs +1≤τ≤t}.

[0083] Among them, t represents the current time step, t obs Indicates the observed time range, τ represents the position and speed characteristics of all traffic participants at time τ, Ε τ Represents lane lines and road boundary features.

[0084] The fast memory module is The difference memory module is Save to or The samples are (X R ',Y R ') and (X D ',Y D '), the training loss function is Then the loss from each data stream S can be expressed as:

[0085]

[0086] Corresponding to the sample (X R ',Y R ') and (X D ',Y D '), let model f θ The output when the sample observation is first input is and The task-free memory replay loss function designed in the present application can be expressed as:

[0087]

[0088] in, and They represent the losses of the fast memory module and the difference memory module respectively, α and β are the weight coefficients of the first loss term and the second loss term respectively, and the total loss function used in this application for training the vehicle behavior prediction model is:

[0089]

[0090] By implementing the above steps 101 to 105, a high prediction accuracy can be always maintained without relying on scene switching information, the model's task-free lifelong learning capability can be realized, and the impact of "catastrophic forgetting" on the vehicle behavior prediction model can be reduced.

[0091] As an optional implementation, the input form of the dynamic and static traffic environment data is a data stream S, and the dynamic and static traffic environment data in the training samples are successively input into the vehicle behavior prediction model. Since the present application is aimed at continuous and changing scenes, the amount of training samples of the dynamic and static traffic environment data collected will change dynamically with the number of scenes encountered by the autonomous driving vehicle, and therefore the total number of samples contained in the data stream S is unknown relative to the vehicle behavior prediction model.

[0092] The present application also provides an application scenario, which applies the above-mentioned vehicle behavior prediction method with task-free lifelong learning capability. Specifically: the vehicle behavior prediction method with task-free lifelong learning capability provided in this embodiment can be applied in an intelligent traffic management system. The intelligent traffic management system includes a data collection link, a data analysis and processing link, and a behavior prediction and response link; the vehicle behavior data enters the data analysis and processing link from the data collection link, and is deeply learned and analyzed by a model with task-free lifelong learning capability to obtain the vehicle's behavior characteristics and trends, and enters the downstream behavior prediction and response link. The vehicle behavior prediction method provided in this embodiment belongs to the core prediction part in the data analysis and processing link. Specifically, in the process of data analysis and processing for vehicle behavior, the task-free lifelong learning model can be relied on to conduct a comprehensive analysis of the vehicle's historical behavior and real-time data, predict the vehicle's future driving trajectory and possible behaviors, and thus provide accurate decision support for traffic management and control.

[0093] Based on the same inventive concept, the embodiment of the present application also provides a vehicle behavior prediction device with task-free lifelong learning capability for realizing the above-mentioned vehicle behavior prediction device. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above-mentioned method, so the specific limitations in one or more embodiments of the vehicle behavior prediction device with task-free lifelong learning capability provided below can refer to the limitations of the vehicle behavior prediction method with task-free lifelong learning capability above, and will not be repeated here.

[0094] In an exemplary embodiment, Figure 2 As shown, a vehicle behavior prediction device with task-free lifelong learning capability is provided, comprising:

[0095] The environmental data acquisition module 201 is used to monitor the traffic environment and obtain dynamic and static traffic environment data of the current time period; the current time period includes the current moment and the first preset number of moments before the current moment; the dynamic and static traffic environment data include: the position information and speed information of all traffic participants in the current traffic environment, and road feature information; traffic participants include target vehicles and vehicles within a preset range of the target vehicle.

[0096] The memory module construction module 202 is used to construct fast memory modules and difference memory modules of a preset number threshold size.

[0097] The first data module 203 is used to store the dynamic and static traffic environment data of the current time period into the fast memory module if the fast memory module has not reached the first saturation state; if the fast memory module reaches the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method; wherein the first saturation state is when the amount of data stored in the fast memory module reaches a preset quantity threshold.

[0098] The second data module 204 is used to store the dynamic and static traffic environment data of the current time period into the difference memory module if the difference memory module has not reached the second saturation state; if the difference memory module reaches the second saturation state, a gradient selection sampling method is used to update the dynamic and static traffic environment data in the difference memory module; wherein the second saturation state is that the amount of data stored in the difference memory module reaches a preset quantity threshold.

[0099] The predicted trajectory acquisition module 205 is used to input the dynamic and static traffic environment data of the current time period in the fast memory module and the difference memory module into the vehicle behavior prediction model, train the vehicle behavior prediction model through the task-free memory playback loss function, and output the predicted trajectory of the target vehicle within the predicted time period; the predicted time period includes a second preset number of moments after the current moment.

[0100] As an optional implementation, the vehicle behavior prediction model structure is as follows: Figure 3 As shown in the figure, the details are as follows: Assuming driving in a continuous and changing scene, the acquired dynamic and static traffic environment data is input into the prediction model. The encoder is used to abstract the dynamic and static traffic environment data in S1 into a hidden representation of the feature: H M , H C , H TV , H S , then input the predictor module and the regularization network module respectively, and finally fuse the outputs of the two modules as the behavior prediction result.

[0101] The encoder uses a layered graph neural network structure. First, the acquired information is sampled into directed spline curves and divided into lane line information, grid coordinate information, and motion trajectory information according to the data type. Among them, the grid coordinate information is converted into a hidden representation H by a multilayer neural network (MultilayerNeuralNetwork, MLP). C ; The lane line information is converted into a hidden representation H through a module combining a graph neural network and an attention mechanism. M After the motion trajectory information passes through a graph neural network module, it is converted into hidden representations H according to the characteristics of the target vehicle and the surrounding vehicles. TV , H S .

[0102] The prediction model proposed in this application uses the past t obs The information of all traffic participants within seconds is used to predict the target vehicle in the next t pre For example, let t obs =2s,t pre =3s.

[0103] Among them, implementing this embodiment, such as Figure 3 As shown in Figure 1, the predictor uses the structure of the interaction module and the decoder. The interaction module is composed of 4 graph convolutional network stacks with a self-attention mechanism added, and its input is H M , H C , H TV , H S , whose output passes through a decoder composed of a two-layer neural network (2-layer MLP) to obtain the first predicted trajectory is a collection of trajectory points over a period of time The trajectory points are represented by the probability of each grid cell in the gridded map.

[0104] Among them, implementing this embodiment, such as Figure 3 As shown in Figure 2, the regularized network uses a two-layer multi-head self-attention layer structure that takes into account the causal relationship of uncertainty quantification. Its input is H M , H C , H TV The first attention mechanism layer passes the map information to the target vehicle, and the second attention mechanism layer generates the probability of each position from the cascade output of the previous layer. Get the second predicted trajectory. is a collection of trajectory points over a period of time The trajectory points are represented by the probability of each grid cell in the gridded map.

[0105] The difference between the regularized network and the predictor network is the input information. The regularized network only inputs the predicted target traffic participant and road feature information, and does not consider the feature information of the traffic participants around the predicted target. The predictor network considers the prediction target, surrounding traffic participants, and road features. The prediction results of the regularized network module and the predictor module are combined to obtain the final behavior prediction result of the entire model.

[0106] The setting of the training loss function takes into account the causal relationship of uncertainty quantification and is expressed as:

[0107]

[0108] in, is the L1 norm, represents the first predicted trajectory, represents the second predicted trajectory, Indicates that Linearly renormalize from 0 to 1.

[0109] The actual trajectory Y of the target vehicle is the set of trajectory points within a period of time {s τ |t≤τ≤t+t pre}, the trajectory point is represented by the probability of each grid in the grid map; l(·) is the loss function selected considering the data characteristics. In order to make the loss function reflect the comparison between the grid probability and the vehicle position coordinates in the actual trajectory, the focal loss is used to construct the loss function:

[0110]

[0111] Among them, the subscript p represents each set of data in the set, and P is the capacity of the set.

[0112] The beneficial effects of the present application are as follows: First, the vehicle behavior prediction method proposed in the present application does not rely on the task labels of the data during training. Most of the existing machine learning models that implement lifelong learning strategies use training data that rely on scene switching information. In the actual application of autonomous driving, different scenes are difficult to identify and define in real time by the autonomous driving system, and scene switching information is difficult to obtain. Therefore, compared with the existing behavior prediction methods, the method disclosed in the present application that does not rely on data task labels has better practicality.

[0113] Secondly, the method disclosed in this application includes two methods of sampling from the data stream and saving it to the memory module: reservoir sampling and sampling based on gradient selection. On the one hand, these sampling methods can make the memory data samples stored in the memory module random, which is suitable for equally likely sampling from data streams of unknown total length; on the other hand, they maximize the differences between data samples and improve the data diversity of the memory module.

[0114] Finally, the method disclosed in the present application can enable the vehicle behavior prediction model to obtain good prediction performance in actual applications when the vehicle re-enters the historical scene without using historical scene data and retraining the model. It has task-free lifelong learning capabilities, always has high prediction accuracy in continuously changing scenes, alleviates catastrophic forgetting, and achieves efficient application of the vehicle behavior prediction model in continuous and changing scenes.

[0115] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store dynamic and static traffic environment data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a vehicle behavior prediction method with task-free lifelong learning capability is implemented.

[0116] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0117] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0118] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0119] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0120] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0121] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0122] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0123] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A vehicle behavior prediction method with task-free lifelong learning capability, characterized in that: The vehicle behavior prediction method with task-free lifelong learning capability includes: The traffic environment is monitored to obtain dynamic and static traffic environment data of the current time period; the current time period includes the current moment and a first preset number of moments before the current moment; the dynamic and static traffic environment data includes: position information and speed information of all traffic participants in the current traffic environment, and road feature information; traffic participants include a target vehicle and vehicles within a preset range of the target vehicle; Construct fast memory modules and difference memory modules with preset number threshold sizes; If the fast memory module has not reached the first saturation state, the dynamic and static traffic environment data of the current time period are stored in the fast memory module; if the fast memory module has reached the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method; wherein the first saturation state is when the amount of data stored in the fast memory module reaches a preset quantity threshold; If the difference memory module has not reached the second saturation state, the dynamic and static traffic environment data of the current time period is stored in the difference memory module; if the difference memory module reaches the second saturation state, the dynamic and static traffic environment data in the difference memory module is updated using a gradient selection sampling method; wherein the second saturation state is when the amount of data stored in the difference memory module reaches a preset quantity threshold; The dynamic and static traffic environment data of the current time period in the fast memory module and the difference memory module are input into the vehicle behavior prediction model, the vehicle behavior prediction model is trained by the task-free memory playback loss function, and the predicted trajectory of the target vehicle in the prediction time period is output; the prediction time period includes a second preset number of moments after the current moment.

2. The vehicle behavior prediction method with task-free lifelong learning capability according to claim 1, characterized in that: If the fast memory module reaches the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method, specifically including: Generate a random number r1 between the interval [0, j-1], and compare the generated random number r1 with a preset number threshold; If the generated random number r1 is less than the preset number threshold, the r1th dynamic and static traffic environment data stored in the fast memory module is replaced with the dynamic and static traffic environment data of the current time period; If the generated random number r1 is greater than or equal to the preset quantity threshold, the fast memory module is not updated.

3. The vehicle behavior prediction method with task-free lifelong learning capability according to claim 1, characterized in that: If the difference memory module reaches the second saturation state, a gradient selection sampling method is used to update the dynamic and static traffic environment data in the difference memory module, specifically including: Calculate the similarity score between the jth dynamic and static traffic environment data in the difference memory module and the preset data set; where, when j=1, the similarity score is set to 0.1; the preset data set is a preset number of dynamic and static traffic environment data randomly selected from the difference memory module except the jth dynamic and static traffic environment data; when j>1, the calculation formula of the similarity score is: Among them, q j represents the similarity score between the jth dynamic and static traffic environment data and the preset data set, g j represents the gradient of the loss function of the jth dynamic and static traffic environment data, g k Represents the loss function gradient of the kth dynamic and static traffic environment data in the preset data set, k = 1, 2, 3, ..., n; k ≠ ​​j, n represents the preset selection number; The interval [0, 1] is The probability size is divided into a preset number threshold molecular interval; among which, P(j) represents the length of the j-th subinterval, M represents the preset number threshold, m=0,1,2,...,M-1; Generate a random number r2 between 0 and 1, and when r2 falls within the j-th subinterval, use the j-th dynamic and static traffic environment data stored in the difference memory module as the dynamic and static traffic environment data to be replaced; Generate a random number r3 between 0 and 1 and compare the random number r3 with q i / (q i +q j ) size; if r3 is smaller than q i / (q i +q j ), replace the dynamic and static traffic environment data to be replaced with the dynamic and static traffic environment data of the current time period; if r3 is greater than or equal to q i / (q i +q j ), the difference memory module is not updated; where q i Indicates the similarity score between the dynamic and static traffic environment data of the current time period and the preset data set.

4. The vehicle behavior prediction method with task-free lifelong learning capability according to claim 1, characterized in that: The calculation formula of the task-free memory replay loss function is: in, represents the task-free memory replay loss function, represents the training loss function, and They represent the fast memory module and the difference memory module respectively. and They represent the losses of the fast memory module and the difference memory module respectively, α and β are the weight coefficients of the first loss term and the second loss term respectively, and f θ represents the vehicle behavior prediction model, X represents the dynamic and static traffic environment data input into the vehicle behavior prediction model, Y represents the actual trajectory of the target vehicle within the prediction time period, and X R ' and X D 'represents the dynamic and static traffic environment data stored in the fast memory module and the difference memory module, respectively, Y R ' indicates X R 'The actual trajectory of the target vehicle during the corresponding prediction period, Y D ' indicates X D 'The actual trajectory of the target vehicle during the corresponding prediction time period, Represents X R 'The predicted trajectory of the target vehicle within the corresponding prediction time period, Represents X R 'The predicted trajectory of the target vehicle within the corresponding prediction time period, represents the mathematical expectation, represents the mathematical expectation of the dynamic and static traffic environment data stored in the fast memory module and the corresponding actual trajectory of the target vehicle, represents the mathematical expectation of the dynamic and static traffic environment data stored in the difference memory module and the corresponding actual trajectory of the target vehicle, represents the mathematical expectation of the dynamic and static traffic environment data stored in the fast memory module and its corresponding target vehicle prediction trajectory, Represents the mathematical expectation of the dynamic and static traffic environment data stored in the difference memory module and its corresponding predicted trajectory of the target vehicle.

5. The vehicle behavior prediction method with task-free lifelong learning capability according to claim 1, characterized in that: The vehicle behavior prediction model includes: an encoder module, a predictor module, and a regularized network module; The encoder module is connected to the predictor module and the regularization network module respectively; The encoder module is used to divide the dynamic and static traffic environment data into grid coordinate information, lane line information, target vehicle motion trajectory information, and motion trajectory information of vehicles within a preset range of the target vehicle; The predictor module is used to obtain a first predicted trajectory of the target vehicle within a predicted time period based on grid coordinate information, lane line information, target vehicle motion trajectory information, and motion trajectory information of vehicles within a preset range of the target vehicle; The regularized network module is used to obtain a second predicted trajectory of the target vehicle within the predicted time period according to the grid coordinate information, the lane line information and the target vehicle motion trajectory information; The first predicted trajectory and the second predicted trajectory are merged to obtain a final predicted trajectory of the target vehicle within the prediction time period.

6. The vehicle behavior prediction method with task-free lifelong learning capability according to claim 4, characterized in that: The calculation formula of the training loss function is: in, is the L1 norm, represents the first predicted trajectory, represents the second predicted trajectory, Indicates that Linearly renormalize from 0 to 1.

7. A vehicle behavior prediction device with task-free lifelong learning capability, characterized in that: The vehicle behavior prediction device with task-free lifelong learning capability comprises: The environment data acquisition module is used to monitor the traffic environment and obtain dynamic and static traffic environment data of the current time period; the current time period includes the current time and the first preset number of time before the current time; the dynamic and static traffic environment data includes: the position information and speed information of all traffic participants in the current traffic environment, and road feature information; traffic participants include target vehicles and vehicles within a preset range of the target vehicle; A building block of a memory module, used to build a fast memory module and a difference memory module of a preset number threshold size; The first data module is used to store the dynamic and static traffic environment data of the current time period into the fast memory module if the fast memory module has not reached the first saturation state; if the fast memory module reaches the first saturation state, the dynamic and static traffic environment data in the fast memory module are updated using a reservoir sampling method; wherein the first saturation state is when the amount of data stored in the fast memory module reaches a preset quantity threshold; A second data module is used to store the dynamic and static traffic environment data of the current time period into the difference memory module if the difference memory module has not reached the second saturation state; if the difference memory module has reached the second saturation state, the dynamic and static traffic environment data in the difference memory module is updated by using a gradient selection sampling method; wherein the second saturation state is when the amount of data stored in the difference memory module reaches a preset quantity threshold; A predicted trajectory acquisition module is used to input the dynamic and static traffic environment data of the current time period in the fast memory module and the difference memory module into the vehicle behavior prediction model, train the vehicle behavior prediction model through the task-free memory playback loss function, and output the predicted trajectory of the target vehicle within the predicted time period; the predicted time period includes a second preset number of moments after the current moment.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the vehicle behavior prediction method with task-free lifelong learning capability as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the vehicle behavior prediction method with task-free lifelong learning capability described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the vehicle behavior prediction method with task-free lifelong learning capability described in any one of claims 1 to 6 is implemented.

Images