A Vehicle Trajectory Prediction Method and System Based on Recurrent Meta-Induced Neural Network

Abstract

The present invention discloses a vehicle trajectory prediction method and system based on a recurrent element-induced neural network, which relates to the field of vehicle trajectory prediction. The method includes: inputting the historical sampling trajectories of a target vehicle and a plurality of corresponding background vehicles into a vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle; the vehicle trajectory prediction model is obtained by training a preset recurrent element-induced neural network based on a training sample set; the preset recurrent element-induced neural network includes an observation prediction sub-network and a target prediction sub-network; the observation prediction sub-network is used to extract the meta-features of the observation prediction sub-task; the target prediction sub-network is used to extract the features of the historical sampling trajectories of the target vehicle and a plurality of corresponding background vehicles to obtain the spatio-temporal interaction features of the vehicle trajectories, and combine the meta-features to predict the future trajectory of the target vehicle. The present invention improves the generalization ability and accuracy of vehicle trajectory prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle trajectory prediction, and particularly to a vehicle trajectory prediction method and system based on a recurrent element-induced neural network. Background Art

[0002] With the continuous development of deep learning technology and the continuous introduction of deep learning methods in the field of autonomous driving, the perception and decision-making of autonomous vehicles have been greatly improved. However, how to improve the interaction ability and driving safety of autonomous driving remains a difficult problem that people focus on. Especially in mixed traffic flows, the high complexity and uncertainty of the traffic environment and vehicle interaction behaviors are the key factors affecting the driving safety of autonomous vehicles, such as the competition and cooperation between human-driven vehicles. Therefore, how autonomous vehicles understand the current interaction scenario, accurately predict the trajectories of surrounding vehicles in the future for a period of time, and then interact with other vehicles safely and effectively has become the research focus of most articles.

[0003] Facing a highly complex and uncertain traffic environment, the problems that need to be solved in vehicle trajectory prediction include: 1) effectively extracting the spatio-temporal social interaction features of vehicle trajectories; 2) improving the generalization ability of trajectory prediction. In the time dimension, there is a certain correlation between the position of a vehicle at the next time step and its position at the previous time step. In the spatial dimension, there is competition and cooperation between vehicles. For example, when the host vehicle encounters a vehicle in another lane changing to the lane where the host vehicle is located, the host vehicle will either brake to give way or accelerate to prevent the other vehicle from changing lanes. Therefore, in both the time and space dimensions, vehicle trajectories are highly complex and non-linear, making it difficult to accurately represent the spatio-temporal interaction features between vehicle trajectories using a precise mathematical model; in addition, due to the high uncertainty of the behaviors of human-driven vehicles, vehicle trajectories have multiple patterns, such as accelerating, decelerating, changing lanes to the left, changing lanes to the right, and maintaining the lane. Simple trajectory prediction methods cannot generalize to all trajectory patterns, resulting in low prediction accuracy.

[0004] Regarding the above two problems, many excellent vehicle trajectory prediction methods have been proposed, which mainly include kinematic model-based and data-driven trajectory prediction methods at present. Among them, kinematic model-based trajectory prediction requires establishing a kinematic model of the prediction object and using Kalman filtering or extended Kalman filtering to predict the trajectory at the next moment. This method can accurately predict the trajectory within 1-2 seconds. However, it cannot extract the spatio-temporal interaction features between relevant vehicles, and thus it is difficult to improve the accuracy of long-term trajectory prediction. The data-driven trajectory prediction method takes into account the high complexity and uncertainty of spatio-temporal interaction between vehicles, and uses a neural network to extract the temporal features and spatial interaction features of vehicle trajectories respectively. It effectively extracts the spatio-temporal interaction features between relevant vehicles and decodes the features, and then predicts the future trajectory of the vehicle. However, this type of method predicts the future trajectory of the target vehicle in a sequence-to-sequence manner. Therefore, to a large extent, it limits the generalization ability of trajectory prediction, and thus it is very difficult to improve the accuracy of trajectory prediction. Summary of the Invention

[0005] The object of the present invention is to provide a vehicle trajectory prediction method and system based on a recurrent element-induced neural network to improve the generalization ability and accuracy of vehicle trajectory prediction.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] In a first aspect, the present invention provides a vehicle trajectory prediction method based on a recurrent element-induced neural network, including:

[0008] Obtaining the historical sampling trajectory of a target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles within a preset time period; the preset time period includes a plurality of sampling moments;

[0009] Inputting the historical sampling trajectory of the target vehicle and the historical sampling trajectories of the plurality of corresponding background vehicles into a vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle; wherein, the vehicle trajectory prediction model is obtained by training a preset recurrent element-induced neural network based on a training sample set; each training sample in the training sample set includes the movement trajectory of the target vehicle, the movement trajectories of the background vehicles corresponding to the target vehicle, and the future trajectory of the target vehicle; the preset recurrent element-induced neural network includes an observation prediction sub-network and a target prediction sub-network; the observation prediction sub-network includes a plurality of observation prediction modules connected in sequence; the target prediction sub-network and each of the observation prediction modules include a first LSTM module and a convolutional social pooling module.

[0010] The observation and prediction sub-network is used to: extract features from the historical sampling trajectories of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at each sampling moment to obtain meta-features; the meta-features include the first vehicle trajectory spatio-temporal interaction feature and the vehicle trajectory prediction feature.

[0011] The target prediction sub-network is used to: extract features from the historical sampling trajectory of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at the target sampling moment to obtain the second vehicle trajectory spatio-temporal interaction feature; then, based on the second vehicle trajectory spatio-temporal interaction feature and the meta-features input by the observation and prediction sub-network, predict the future trajectory of the target vehicle.

[0012] Optionally, the observation and prediction module further includes a first splicing module and a second LSTM module.

[0013] In the observation and prediction module, the input end of the first LSTM module is used to input the historical sampling trajectory of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at any sampling moment; the first LSTM module is used to extract the historical trajectory features of all vehicles and the vehicle trajectory prediction feature of the target vehicle.

[0014] The first output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module; the second output end of the first LSTM module is used to input the vehicle trajectory prediction feature into the first splicing module.

[0015] The convolutional social pooling module is used to extract features from the historical trajectory features of all vehicles to obtain the first vehicle trajectory spatio-temporal interaction feature, and then input the first vehicle trajectory spatio-temporal interaction feature into the first splicing module.

[0016] The first splicing module is used to connect the vehicle trajectory prediction feature and the first vehicle trajectory spatio-temporal interaction feature to obtain the vehicle trajectory comprehensive feature.

[0017] The second LSTM module is used to extract features from the vehicle trajectory comprehensive feature to obtain meta-features, and then input the meta-features into the second LSTM module of the adjacent next observation and prediction module for update.

[0018] Optionally, the target prediction sub-network further includes a second splicing module and a third LSTM module.

[0019] In the target prediction sub-network, the input end of the first LSTM module is used to input the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles at the target sampling moment; the first LSTM module is used to extract the historical trajectory features of all vehicles;

[0020] The output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module; the convolutional social pooling module is used to extract features from the historical trajectory features of all vehicles to obtain the second vehicle trajectory spatio-temporal interaction features, and then input the second vehicle trajectory spatio-temporal interaction features into the second splicing module;

[0021] The second splicing module is also connected to the second LSTM module in the observation prediction module; the second splicing module is used to connect the meta-features input by the second LSTM module with the second vehicle trajectory spatio-temporal interaction features to obtain the target features;

[0022] The third LSTM module is used to extract features from the target features to obtain the target predicted future trajectory.

[0023] Optionally, the convolutional social pooling module includes a convolutional neural network sub-module and a pooling sub-module connected in sequence.

[0024] Optionally, the construction process of the training sample set specifically includes:

[0025] Obtain the movement trajectories of multiple vehicles on a preset road;

[0026] Based on the movement trajectories of the multiple vehicles, determine the target sample vehicle, a plurality of corresponding background sample vehicles of the target sample vehicle, the movement trajectory of the target sample vehicle within a preset sample time, and the movement trajectory of each background sample vehicle within the preset sample time;

[0027] Based on the movement trajectories of the multiple vehicles, determine the future trajectory of the target sample vehicle relative to the preset sample time; the movement trajectory of the target sample vehicle within the preset sample time, the movement trajectory of each background sample vehicle within the preset sample time, and the future trajectory of the target sample vehicle relative to the preset sample time constitute a training sample;

[0028] Multiple such training samples constitute a training sample set.

[0029] Optionally, the determination process of the multiple background sample vehicles corresponding to the target sample vehicle specifically includes:

[0030] Establish a coordinate system with the target sample vehicle as the origin;

[0031] Obtain the inter-vehicle distance between the target sample vehicle and any vehicle in the same lane or adjacent lanes;

[0032] When the inter-vehicle distance is within a preset threshold range, mark the vehicle corresponding to the inter-vehicle distance as a background sample vehicle;

[0033] When the inter-vehicle distance is not within the preset threshold range, discard the vehicle corresponding to the inter-vehicle distance.

[0034] In a second aspect, the present invention provides a vehicle trajectory prediction system based on a recurrent element-induced neural network, including:

[0035] A vehicle data acquisition subsystem for obtaining the historical sampling trajectories of a target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles within a preset time period; the preset time period includes a plurality of sampling moments;

[0036] A vehicle trajectory prediction subsystem for inputting the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles into a vehicle trajectory prediction model to obtain a predicted trajectory of the target vehicle; wherein, the vehicle trajectory prediction model is obtained by training a preset recurrent element-induced neural network based on a training sample set; each training sample in the training sample set includes the motion trajectory of the target vehicle, the motion trajectory of the background vehicle corresponding to the target vehicle, and the future trajectory of the target vehicle; the preset recurrent element-induced neural network includes an observation prediction sub-network and a target prediction sub-network; the observation prediction sub-network includes a plurality of sequentially connected observation prediction modules; the target prediction sub-network and each of the observation prediction modules include a first LSTM module and a convolutional social pooling module;

[0037] The observation prediction sub-network is used for: extracting features from the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles at each sampling moment to obtain meta-features; the meta-features include a first vehicle trajectory spatio-temporal interaction feature and a vehicle trajectory prediction feature;

[0038] The target prediction sub-network is used for: extracting features from the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles at a target sampling moment to obtain a second vehicle trajectory spatio-temporal interaction feature; and then predicting the future trajectory of the target vehicle based on the second vehicle trajectory spatio-temporal interaction feature and the meta-features input by the observation prediction sub-network.

[0039] Optionally, the observation prediction module further includes a first splicing module and a second LSTM module;

[0040] In the observation prediction module, the input end of the first LSTM module is used to input the historical sampling trajectory of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at any sampling time; the first LSTM module is used to extract the historical trajectory features of all vehicles and the vehicle trajectory prediction features of the target vehicle;

[0041] The first output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module; the second output end of the first LSTM module is used to input the vehicle trajectory prediction features into the first splicing module;

[0042] The convolutional social pooling module is used to extract the historical trajectory features of all vehicles to obtain the first vehicle trajectory spatiotemporal interaction features, and then input the first vehicle trajectory spatiotemporal interaction features into the first splicing module;

[0043] The first splicing module is used to connect the vehicle trajectory prediction feature with the first vehicle trajectory spatiotemporal interaction feature to obtain a vehicle trajectory comprehensive feature;

[0044] The second LSTM module is used to extract the comprehensive features of the vehicle trajectory to obtain meta-features, and then input the meta-features into the second LSTM module of the adjacent next observation prediction module for updating.

[0045] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0046] The present invention discloses a vehicle trajectory prediction method and system based on a recurrent element induced neural network, which obtains the historical sampling trajectory of a target vehicle and the corresponding historical sampling trajectory of multiple background vehicles within a preset time period, thereby obtaining stand-by data with high temporal sequence. The mapping of the vehicle's historical trajectory to the future trajectory is regarded as a task. For the task set of vehicle trajectory prediction, a preset recurrent element induced neural network can be constructed based on an LSTM (Long Short-Term Memory) network and convolutional social pooling, so as to effectively extract the spatiotemporal interaction features of the input trajectory and improve the model's representation ability of the spatiotemporal interaction features of the input trajectory. Furthermore, the preset recurrent element induced neural network includes an observation prediction subnetwork and a target prediction subnetwork. By optimizing the model parameters of the two subnetworks, the future trajectory of the target vehicle can be adaptively predicted based on the historical trajectory, thereby improving the generalization ability and accuracy of trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0048] Figure 1 It is a schematic flow chart of the vehicle trajectory prediction method based on the recurrent element-induced neural network of the present invention;

[0049] Figure 2 It is a schematic structural diagram of the preset recurrent element-induced neural network of the present invention;

[0050] Figure 3 It is a schematic structural diagram of the vehicle trajectory prediction system based on the recurrent element-induced neural network of the present invention. Specific embodiments

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0052] The present invention provides a vehicle trajectory prediction method and system based on a recurrent element-induced neural network, which uses a new paradigm based on the recurrent element-induced neural network to establish a trajectory prediction model, and uses the correlation between trajectory prediction tasks to predict the future trajectory of the target vehicle in the target task, improving the generalization ability of the estimation prediction model for different driving behavior patterns and the accuracy of the prediction results; among them, a spatio-temporal interaction feature extraction module based on LSTM and Convolutional social pooling effectively extracts the interaction features between vehicle trajectories, which is more in line with the actual traffic conditions.

[0053] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will further describe the present invention in detail in conjunction with the accompanying drawings and specific embodiments.

[0054] Embodiment 1

[0055] Different from the sequence-to-sequence trajectory prediction, the vehicle trajectory prediction method based on the recurrent element-induced neural network of the present invention uses the observed trajectory prediction subtasks and the historical trajectories of the target task to predict the probability distribution of the future trajectory.

[0056] Assume that the historical trajectories of all relevant vehicles (including the target vehicle and the corresponding multiple background vehicles) are:

[0057] X = (X1, X2, X3, ..., X n ), X ∈ X

[0058] represents the trajectory sequence of the i-th vehicle in the past h time, X represents the historical trajectory space, and n represents the number of relevant vehicles; represents the lateral displacement and longitudinal displacement of the i-th vehicle at the past j moment.

[0059] The future trajectory of the target vehicle is Y = (y t , y t+1 , y t+2 , ..., y t+T ), Y ∈ Y, Y represents the future trajectory space of the target vehicle, and T represents the time length of the future trajectory. y t+j ∈ R 2 , 0 ≤ j ≤ T represents the lateral displacement and longitudinal displacement of the target vehicle at the future j moment.

[0060] Based on the above historical trajectory and future trajectory, the observed trajectory prediction subtask can be obtained as:

[0061] O = (O1, O2, O3, ..., O m )

[0062] The observed trajectory prediction subtask has a high temporal order, and O i ∈ O represents the pair of the historical trajectories of all relevant vehicles and the future trajectory of the target vehicle in the i-th subtask, that is:

[0063] O i = (X, Y), X ∈ X, Y ∈ Y.

[0064] For example, any trajectory prediction subtask can be: extracting the lateral positions and longitudinal positions of all relevant vehicles (including the target vehicle and corresponding multiple background vehicles) in the past 3s at a sampling frequency of 5HZ, that is, the historical trajectories of all relevant vehicles, and the lateral positions and longitudinal positions of the prediction object in the successive 5s, that is, the future trajectory of the target vehicle. Specifically, the sampled historical time and future prediction time can be adjusted by relevant personnel according to actual needs. And the sampling frequency can determine the sampling time interval k, where the sampling time interval k refers to the time difference between O i and O i-1 .

[0065] In summary, the trajectory prediction formula described by the vehicle trajectory prediction method based on the recurrent element-induced neural network is as follows:

[0066]

[0067] where ξ is a parameter of the probability distribution P ξ ; is the prediction result, representing the trajectory sequence of the target vehicle within the future time T; X Tar represents the historical trajectories of all relevant vehicles in the target task.

[0068] To implement the prediction based on the above trajectory prediction formula, the present invention proposes a vehicle trajectory prediction method based on a recurrent element-induced neural network, as Figure 1 shown, the method includes:

[0069] Step 100, obtaining the historical sampled trajectories of the target vehicle and the historical sampled trajectories of a plurality of corresponding background vehicles within a preset time period; the preset time period includes a plurality of sampling moments. Specifically, the preset time period may be 3 seconds.

[0070] Step 200, inputting the historical sampled trajectories of the target vehicle and the historical sampled trajectories of a plurality of corresponding background vehicles into a vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle; wherein, the vehicle trajectory prediction model is obtained by training a preset recurrent element-induced neural network based on a training sample set.

[0071] Each training sample in the training sample set includes the motion trajectory of the target vehicle, the motion trajectories of the background vehicles corresponding to the target vehicle, and the future trajectory of the target vehicle. Specifically, the construction process of the training sample set specifically includes:

[0072] 1) Obtaining the motion trajectories of a plurality of vehicles on a preset road. In a specific example, obtaining the I-80 and US-101 data in the NGSIM dataset.

[0073] 2) Based on the motion trajectories of the plurality of vehicles, determining a target sample vehicle, a plurality of corresponding background sample vehicles of the target sample vehicle, the motion trajectory of the target sample vehicle within a preset sample time, and the motion trajectories of each background sample vehicle within the preset sample time.

[0074] Specifically, first, select multiple target sample vehicles from the obtained data. For each target sample vehicle, establish a coordinate system with the target sample vehicle as the origin; obtain the inter-vehicle distance between the target sample vehicle and any vehicle in the same lane or adjacent lanes; when the inter-vehicle distance is within a preset threshold range, mark the vehicle corresponding to the inter-vehicle distance as a background sample vehicle; when the inter-vehicle distance is not within the preset threshold range, discard the vehicle corresponding to the inter-vehicle distance, so as to obtain multiple background vehicles. Convert the lateral position coordinates and longitudinal position coordinates of all background vehicles to the coordinate system with the target vehicle as the coordinate origin, and represent the coordinate positions of the background vehicles after coordinate conversion in the form of the historical trajectory space X mentioned above, so as to facilitate subsequent model training and model prediction.

[0075] 3) Based on the motion trajectories of the multiple vehicles, determine the future trajectory of the target sample vehicle relative to a preset sample time. Specifically, from the motion trajectories of the multiple vehicles, intercept the trajectory of the target sample vehicle within a certain time after the end of the preset sample time as its future trajectory relative to the preset sample time. The certain time after the end of the preset sample time can be 5s later.

[0076] The motion trajectory of the target sample vehicle within the preset sample time, the motion trajectory of each background sample vehicle within the preset sample time, and the future trajectory of the target sample vehicle relative to the preset sample time constitute a training sample. Specifically, take the motion trajectories of the target sample vehicle and the corresponding multiple background sample vehicles within the preset sample time as the historical trajectories of the relevant vehicles, and take the future trajectory of the target sample vehicle relative to the preset sample time as the label corresponding to the historical trajectories of the relevant vehicles. These two constitute a subtask in task O.

[0077] 4) Multiple such training samples constitute a training sample set.

[0078] As Figure 2 shown, the preset recurrent element induction neural network includes an observation prediction sub-network (corresponding to Figure 2 the Observations subtasks in Figure 2 ) and a target prediction sub-network (corresponding to

[0079] the Target task in Figure 2 ); the observation prediction sub-network includes multiple sequentially connected observation prediction modules; the target prediction sub-network and each of the observation prediction modules include a first LSTM module and a convolutional social pooling module.

[0079] The observation prediction sub-network is used to: extract features from the historical sampling trajectories of the target vehicle and the corresponding multiple background vehicles at each sampling moment to obtain meta-features; the meta-features include a first vehicle trajectory spatio-temporal interaction feature and a vehicle trajectory prediction feature.

[0080] The target prediction sub-network is used to: extract features from the historical sampling trajectories of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at the target sampling moment to obtain the second vehicle trajectory spatio-temporal interaction features; then, based on the second vehicle trajectory spatio-temporal interaction features and the meta-features input by the observation prediction sub-network, determine the future trajectory of the target. Among them, the target sampling moment is generally the end moment of a preset time period, and the vehicle trajectory within a certain time after this moment is predicted based on the end moment of the preset time period.

[0081] Furthermore, the observation prediction module further includes a first concatenation module (corresponding to Figure 2 concatenate in it) and a second LSTM module. In the observation prediction module, the input end of the first LSTM module is used to input the historical sampling trajectories of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at any sampling moment. The first LSTM module is used to extract the historical trajectory features of all vehicles and the vehicle trajectory prediction features of the target vehicle, and store the historical trajectory features of all vehicles in a grid data structure.

[0082] The first output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module. The second output end of the first LSTM module is used to input the vehicle trajectory prediction features into the first concatenation module.

[0083] The convolutional social pooling module is used to extract features from the historical trajectory features of all vehicles to obtain the first vehicle trajectory spatio-temporal interaction features, and then input the first vehicle trajectory spatio-temporal interaction features into the first concatenation module; the convolutional social pooling module includes a convolutional neural network sub-module (corresponding to Figure 2 CNN in it) and a pooling sub-module (corresponding to Figure 2 MP in it, that is, maxpooling, the max pooling layer).

[0084] The first concatenation module is used to concatenate the vehicle trajectory prediction features and the first vehicle trajectory spatio-temporal interaction features to obtain the vehicle trajectory comprehensive features.

[0085] The second LSTM module is used to extract features from the vehicle trajectory comprehensive features to obtain meta-features, and then input the meta-features into the second LSTM module of the adjacent next observation prediction module for updating.

[0086] Such as Figure 2As shown, there are n observation and prediction modules in total. When n = 1, the second LSTM module only needs to receive the comprehensive vehicle trajectory features input by the first splicing module. When n > 1, the second LSTM module needs to receive the comprehensive vehicle trajectory features input by the corresponding first splicing module and the meta-features input by the second LSTM module of the (n - 1)th observation and prediction module, so as to perform feature extraction in the second LSTM module of the nth observation and prediction module and output the updated meta-features. When n reaches the preset maximum value, the second LSTM module of the nth observation and prediction module outputs the updated meta-features (corresponding to Figure 2 r in

[0087] ), and then inputs it into the target prediction sub-network. The target prediction sub-network further includes a second splicing module and a third LSTM module. In the target prediction sub-network, the input end of the first LSTM module is used to input the historical sampling trajectories of the target vehicle and the corresponding historical sampling trajectories of multiple background vehicles at the target sampling moment; the first LSTM module is used to extract the historical trajectory features of all vehicles.

[0088] The output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module; the convolutional social pooling module is used to extract features from the historical trajectory features of all vehicles to obtain the second vehicle trajectory spatio-temporal interaction features, and then input the second vehicle trajectory spatio-temporal interaction features into the second splicing module.

[0089] The second splicing module is also connected to the second LSTM module in the observation and prediction module; the second splicing module is used to connect the meta-features input by the second LSTM module with the second vehicle trajectory spatio-temporal interaction features to obtain the target features.

[0090] The third LSTM module is used to extract features from the target features to obtain the predicted future trajectory of the target. In a specific embodiment, the third LSTM module outputs the trajectory of the target vehicle within the next 5 seconds.

[0091] Based on the training sample set, the preset recurrent meta-induced neural network constructed above is trained and tested to obtain the preset recurrent meta-induced neural network with the optimal effect; the preset recurrent meta-induced neural network with the optimal effect is used as the vehicle trajectory prediction model to participate in the actual vehicle trajectory prediction, so as to adaptively predict the probability distribution of the future trajectory according to the features of the observation sub-task and the historical trajectories in the target task.

[0092] Embodiment 2

[0093] As Figure 3As shown in the figure, in order to execute the method corresponding to the first embodiment above to achieve the corresponding functions and technical effects, this embodiment also provides a vehicle trajectory prediction system based on a recurrent element-induced neural network. The system includes:

[0094] A vehicle data acquisition subsystem 101, configured to acquire the historical sampling trajectories of a target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles within a preset time period; the preset time period includes a plurality of sampling moments.

[0095] A vehicle trajectory prediction subsystem 201, configured to input the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles into a vehicle trajectory prediction model to obtain a predicted trajectory of the target vehicle; wherein, the vehicle trajectory prediction model is obtained by training a preset recurrent element-induced neural network based on a training sample set; each training sample in the training sample set includes the motion trajectory of the target vehicle, the motion trajectory of the background vehicle corresponding to the target vehicle, and the future trajectory of the target vehicle; the preset recurrent element-induced neural network includes an observation prediction sub-network and a target prediction sub-network; the observation prediction sub-network includes a plurality of sequentially connected observation prediction modules; the target prediction sub-network and each of the observation prediction modules include a first LSTM module and a convolutional social pooling module.

[0096] The observation prediction sub-network is configured to: extract features from the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles at each sampling moment to obtain meta-features; the meta-features include a first vehicle trajectory spatio-temporal interaction feature and a vehicle trajectory prediction feature.

[0097] The target prediction sub-network is configured to: extract features from the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles at a target sampling moment to obtain a second vehicle trajectory spatio-temporal interaction feature; and then predict the future trajectory of the target vehicle based on the second vehicle trajectory spatio-temporal interaction feature and the meta-features input by the observation prediction sub-network.

[0098] Specifically, the observation and prediction module further includes a first splicing module and a second LSTM module; in the observation and prediction module, the input end of the first LSTM module is used to input the historical sampling trajectory of the target vehicle and the historical sampling trajectories of the corresponding multiple background vehicles at any sampling moment; the first LSTM module is used to extract the historical trajectory features of all vehicles and the vehicle trajectory prediction features of the target vehicle; the first output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module; the second output end of the first LSTM module is used to input the vehicle trajectory prediction features into the first splicing module; the convolutional social pooling module is used to extract features from the historical trajectory features of all vehicles to obtain the first vehicle trajectory spatio-temporal interaction feature, and then input the first vehicle trajectory spatio-temporal interaction feature into the first splicing module; the first splicing module is used to connect the vehicle trajectory prediction feature and the first vehicle trajectory spatio-temporal interaction feature to obtain the vehicle trajectory comprehensive feature; the second LSTM module is used to extract features from the vehicle trajectory comprehensive feature to obtain the meta-feature, and then input the meta-feature into the second LSTM module of the adjacent next observation and prediction module for update.

[0099] The various embodiments in this specification are described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple. For the relevant parts, please refer to the description in the method section.

[0100] Specific examples are used in this article to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A vehicle trajectory prediction method based on a recurrent meta-induced neural network, characterized in that, The method includes: Obtaining the historical sampling trajectories of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles within a preset time period; the preset time period includes a plurality of sampling moments; Inputting the historical sampling trajectories of the target vehicle and the historical sampling trajectories of the corresponding plurality of background vehicles into a vehicle trajectory prediction model to obtain a predicted trajectory of the target vehicle; wherein, the vehicle trajectory prediction model is obtained by training a preset recurrent element induction neural network based on a training sample set; each training sample in the training sample set includes the motion trajectory of the target vehicle, the motion trajectory of the background vehicle corresponding to the target vehicle, and the future trajectory of the target vehicle; the preset recurrent element induction neural network includes an observation prediction sub-network and a target prediction sub-network; the observation prediction sub-network includes a plurality of sequentially connected observation prediction modules; the target prediction sub-network and each of the observation prediction modules include a first LSTM module and a convolutional social pooling module; The observation prediction sub-network is configured to: extract features from the historical sampling trajectories of the target vehicle and the historical sampling trajectories of the corresponding plurality of background vehicles at each sampling moment to obtain meta-features; the meta-features include a first vehicle trajectory spatio-temporal interaction feature and a vehicle trajectory prediction feature; The target prediction sub-network is configured to: extract features from the historical sampling trajectories of the target vehicle and the historical sampling trajectories of the corresponding plurality of background vehicles at a target sampling moment to obtain a second vehicle trajectory spatio-temporal interaction feature; and then predict the future trajectory of the target vehicle based on the second vehicle trajectory spatio-temporal interaction feature and the meta-features input by the observation prediction sub-network; The observation prediction module further includes a first splicing module and a second LSTM module; In the observation prediction module, the input end of the first LSTM module is configured to input the historical sampling trajectories of the target vehicle and the historical sampling trajectories of the corresponding plurality of background vehicles at any sampling moment; the first LSTM module is configured to extract the historical trajectory features of all vehicles and the vehicle trajectory prediction feature of the target vehicle; The first output end of the first LSTM module is configured to input the historical trajectory features of all vehicles to the convolutional social pooling module; the second output end of the first LSTM module is configured to input the vehicle trajectory prediction feature to the first splicing module; The convolutional social pooling module is configured to extract features from the historical trajectory features of all vehicles to obtain a first vehicle trajectory spatio-temporal interaction feature, and then input the first vehicle trajectory spatio-temporal interaction feature to the first splicing module; The first splicing module is configured to connect the vehicle trajectory prediction feature and the first vehicle trajectory spatio-temporal interaction feature to obtain a vehicle trajectory comprehensive feature; The second LSTM module is configured to extract features from the vehicle trajectory comprehensive feature to obtain meta-features, and then input the meta-features to the second LSTM module of the adjacent next observation prediction module for update.

2. The vehicle trajectory prediction method based on a recurrent element-induced neural network according to claim 1, characterized in that The target prediction sub-network further includes a second splicing module and a third LSTM module; In the target prediction sub-network, the input end of the first LSTM module is used to input the historical sampling trajectory of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles at the target sampling moment; the first LSTM module is used to extract the historical trajectory features of all vehicles; The output end of the first LSTM module is used to input the historical trajectory features of all vehicles into the convolutional social pooling module; the convolutional social pooling module is used to extract features from the historical trajectory features of all vehicles to obtain the second vehicle trajectory spatio-temporal interaction features, and then input the second vehicle trajectory spatio-temporal interaction features into the second splicing module; The second splicing module is also connected to the second LSTM module in the observation prediction module; the second splicing module is used to connect the meta-features input by the second LSTM module with the second vehicle trajectory spatio-temporal interaction features to obtain the target features; The third LSTM module is used to extract features from the target features to obtain the target predicted future trajectory.

3. The vehicle trajectory prediction method based on a recurrent element-induced neural network according to claim 1, wherein The convolutional social pooling module includes a convolutional neural network sub-module and a pooling sub-module connected in sequence.

4. The vehicle trajectory prediction method based on a recurrent element-induced neural network according to claim 1, wherein The construction process of the training sample set specifically includes: Obtain the movement trajectories of multiple vehicles on a preset road; Based on the movement trajectories of the multiple vehicles, determine the target sample vehicle, a plurality of corresponding background sample vehicles of the target sample vehicle, the movement trajectory of the target sample vehicle within the preset sample time, and the movement trajectory of each background sample vehicle within the preset sample time; Based on the movement trajectories of the multiple vehicles, determine the future trajectory of the target sample vehicle relative to the preset sample time; the movement trajectory of the target sample vehicle within the preset sample time, the movement trajectory of each background sample vehicle within the preset sample time, and the future trajectory of the target sample vehicle relative to the preset sample time constitute a training sample; Multiple such training samples constitute a training sample set.

5. The vehicle trajectory prediction method based on a cyclic element-induced neural network according to claim 4, wherein The determination process of the plurality of background sample vehicles corresponding to the target sample vehicle specifically includes: Establish a coordinate system with the target sample vehicle as the origin; Obtain the inter-vehicle distance between the target sample vehicle and any vehicle in the same lane or an adjacent lane; When the inter-vehicle distance is within the preset threshold range, mark the vehicle corresponding to the inter-vehicle distance as a background sample vehicle; When the inter-vehicle distance is not within the preset threshold range, discard the vehicle corresponding to the inter-vehicle distance.

6. A vehicle trajectory prediction system based on a recurrent meta-induced neural network, characterized in that, The system includes: A vehicle data acquisition subsystem, configured to acquire the historical sampling trajectory of the target vehicle and the historical sampling trajectories of a plurality of corresponding background vehicles within a preset time period; the preset time period includes a plurality of sampling moments; A vehicle trajectory prediction subsystem is configured to input the historical sampled trajectories of the target vehicle and the corresponding historical sampled trajectories of multiple background vehicles into a vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle. The vehicle trajectory prediction model is obtained by training a preset recurrent element induction neural network based on a training sample set. Each training sample in the training sample set includes the motion trajectory of the target vehicle, the motion trajectory of the background vehicle corresponding to the target vehicle, and the future trajectory of the target vehicle. The preset recurrent element induction neural network includes an observation prediction subnetwork and a target prediction subnetwork. The observation prediction subnetwork includes a plurality of sequentially connected observation prediction modules. The target prediction subnetwork and each of the observation prediction modules include a first LSTM module and a convolutional social pooling module. The observation prediction subnetwork is configured to: extract features from the historical sampled trajectories of the target vehicle and the corresponding historical sampled trajectories of multiple background vehicles at each sampling moment to obtain meta-features. The meta-features include a first vehicle trajectory spatio-temporal interaction feature and a vehicle trajectory prediction feature. The target prediction subnetwork is configured to: extract features from the historical sampled trajectories of the target vehicle and the corresponding historical sampled trajectories of multiple background vehicles at the target sampling moment to obtain a second vehicle trajectory spatio-temporal interaction feature. Then, based on the second vehicle trajectory spatio-temporal interaction feature and the meta-features input by the observation prediction subnetwork, predict the future trajectory of the target vehicle. The observation prediction module further includes a first splicing module and a second LSTM module. In the observation prediction module, the input end of the first LSTM module is configured to input the historical sampled trajectories of the target vehicle and the corresponding historical sampled trajectories of multiple background vehicles at any sampling moment. The first LSTM module is configured to extract the historical trajectory features of all vehicles and the vehicle trajectory prediction feature of the target vehicle. The first output end of the first LSTM module is configured to input the historical trajectory features of all vehicles into the convolutional social pooling module. The second output end of the first LSTM module is configured to input the vehicle trajectory prediction feature into the first splicing module. The convolutional social pooling module is configured to extract features from the historical trajectory features of all vehicles to obtain a first vehicle trajectory spatio-temporal interaction feature, and then input the first vehicle trajectory spatio-temporal interaction feature into the first splicing module. The first splicing module is configured to connect the vehicle trajectory prediction feature and the first vehicle trajectory spatio-temporal interaction feature to obtain a vehicle trajectory comprehensive feature. The second LSTM module is configured to extract features from the vehicle trajectory comprehensive feature to obtain meta-features, and then input the meta-features into the second LSTM module of the adjacent next observation prediction module for update.

Images