Vehicle trajectory prediction method, device, computer, readable storage medium and motor vehicle

Abstract

The present invention discloses a vehicle trajectory prediction method, device, computer, readable storage medium and motor vehicle, which relate to the field of autonomous driving technology and are used to predict the vehicle trajectory during autonomous driving, comprising the following steps: obtaining interaction features of a target vehicle and its surrounding vehicles based on historical motion information; obtaining scene features of the target vehicle and its adjacent lanes based on lane centerline information; splicing the interaction features, scene features and time series features of the target vehicle to form fusion features; calculating future interaction features based on the fusion features; splicing and decoding the fusion features and future interaction features to form a predicted trajectory of the target vehicle. The method provided by the present invention accurately extracts the interaction features of surrounding vehicles, making the trajectory prediction results of surrounding vehicles more accurate, thereby ensuring that the autonomous driving vehicle completes reasonable trajectory planning and controls the vehicle to travel smoothly.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a vehicle trajectory prediction method, device, computer, readable storage medium, and motor vehicle. Background Art

[0002] When a driver is driving a motor vehicle, he or she will pay attention to the surrounding lane information, as well as the movement status of other motor vehicles in adjacent lanes, and predict the driving paths of motor vehicles within a certain range to avoid collisions. If an autonomous vehicle wants to achieve complete unmanned driving, it also needs to judge the vehicles around it, predict the trajectories of the surrounding vehicles in the future, and plan the driving trajectory of the autonomous vehicle in the future based on the prediction results. It completes the planning and control of the autonomous vehicle trajectory to ensure that there will be no collisions or conflicts with surrounding vehicles. The existing technology mainly uses the interaction information between the target vehicle to be predicted and the surrounding vehicles, as well as the interaction information between the target vehicle to be predicted and the surrounding environment to encode features, mainly considering historical interaction information and scene interaction information. However, the future trajectory information of the target vehicle to be predicted is also an interaction feature that needs to be considered. The existing technology does not take this into account, which affects the rationality of the trajectory calculation of the autonomous vehicle. Summary of the Invention

[0003] To solve the aforementioned problems, the present invention provides a vehicle trajectory prediction method that accurately extracts the interaction features of surrounding vehicles, making the trajectory prediction results of surrounding vehicles more accurate, thereby ensuring that the autonomous driving vehicle completes reasonable trajectory planning and controls the vehicle to travel smoothly.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A vehicle trajectory prediction method is used to predict the vehicle trajectory during automatic driving of the vehicle, and the vehicle trajectory prediction method comprises the following steps:

[0006] Obtain the interaction features between the target vehicle and its surrounding vehicles based on their historical motion information;

[0007] The scene features of both the target vehicle and its adjacent lanes are obtained based on their historical motion information.

[0008] Combine the interaction features, scene features, and temporal features of the target vehicle to form fusion features;

[0009] Calculate future interaction features based on fused features;

[0010] The fused features and future interaction features are concatenated and decoded to form the predicted trajectory of the target vehicle.

[0011] Optionally, obtaining interaction features of the target vehicle and its surrounding vehicles based on historical motion information thereof includes the following steps:

[0012] Obtain the historical motion information of the target vehicle and its surrounding vehicles respectively;

[0013] The temporal convolutional network is used to encode the historical motion information of the target vehicle and obtain the temporal characteristics of the target vehicle;

[0014] Use the temporal convolutional network to encode the historical motion information of surrounding vehicles and obtain the temporal features of surrounding vehicles;

[0015] Use one of several fully connected layers to encode the temporal features of the target vehicle to obtain the query feature Q0;

[0016] The other two layers in the fully connected layers are used to encode the temporal features of the surrounding vehicles to obtain the Key feature K0 and the Value feature V0;

[0017] The query feature Q0, key feature K0, and value feature V0 are fused according to the following formula to obtain the interactive feature:

[0018]

[0019] Among them, f interaction is the interaction feature, d o is the Key feature K o The feature dimension of ; T represents the matrix transpose.

[0020] Optionally, obtaining scene features of the target vehicle and its adjacent lanes based on their historical motion information and lane centerline information includes the following steps:

[0021] Obtain lane centerline information of adjacent lanes of the target vehicle;

[0022] Use the PointNet network to encode the lane centerline information and obtain lane line features;

[0023] Use one of several fully connected layers to encode the temporal features of the target vehicle and obtain the query feature Q1

[0024] The other two layers in the fully connected layers are used to encode the lane features respectively to obtain the Key feature K1 and the Value feature V1;

[0025] The query feature Q1, key feature K1, and value feature V1 are fused according to the following formula to obtain the scene feature:

[0026]

[0027] Among them, f scene is the scene feature; d l is the Key feature K l The feature dimension of ; T represents the matrix transpose.

[0028] Optionally, the interaction features, scene features, and target vehicle temporal features are concatenated according to the following formula:

[0029] f concat =[f target ;f interaction ;f scene ]

[0030] Among them, f concat For fusion features.

[0031] Optionally, calculating future interaction features based on fused features includes the following steps:

[0032] Several fully connected layers are used to encode the fused features and obtain the initial predicted future trajectory;

[0033] The temporal convolutional network is used to extract the temporal features of the preliminary predicted future trajectory to obtain the temporal features of the preliminary predicted future trajectory;

[0034] Use one of several fully connected layers to encode the fusion features and obtain the query feature Q f

[0035] The other two layers in the fully connected layers are used to encode the temporal features of the predicted preliminary future trajectory to obtain the Key feature K f and Value feature V f ;

[0036] According to the following formula, the Query feature Q f , Key feature K f 、Value feature V f Fusion is performed to obtain future interaction features:

[0037]

[0038] Among them, f future_interact For future interactive features, d f is the Key feature K f The feature dimension of ; T represents the matrix transpose.

[0039] Optionally, concatenating and decoding the fused features and the future interaction features to form a predicted trajectory of the target vehicle includes the following steps:

[0040] Concatenate the fusion features and future interaction features to form encoding features;

[0041] Several fully connected layers are used to decode the encoded features to form the predicted trajectory of the target vehicle.

[0042] Optionally, the fusion features and future interaction features are concatenated according to the following formula:

[0043] f encoder =[f concat ;f uture_interact]

[0044] Among them, f encoder is the coding feature.

[0045] The present invention has the following beneficial effects:

[0046] The technical solution provided by the present invention combines the temporal convolutional neural network and the cross-channel attention mechanism, takes into account the interaction information between vehicles, accurately extracts the interaction features of surrounding vehicles, and makes the trajectory prediction results of surrounding vehicles more accurate, thereby ensuring that the autonomous driving vehicle completes reasonable trajectory planning and controls the vehicle to drive smoothly.

[0047] Correspondingly, the present invention further provides a vehicle trajectory prediction device for predicting a vehicle trajectory during automatic driving of the vehicle, the vehicle trajectory prediction device comprising:

[0048] A fully connected neural network is used to obtain interaction features between the target vehicle and its surrounding vehicles based on their historical motion information; and to obtain scene features of the target vehicle and its adjacent lanes based on their historical motion information and lane centerline information.

[0049] A splicing module is used to splice the interaction features, scene features and the time series features of the target vehicle to form a fusion feature; and to splice the fusion feature and the future interaction features;

[0050] The fully connected neural network is further used to calculate future interaction features based on the fused features; and to decode the fused features and the future interaction features after splicing them together to form a predicted trajectory of the target vehicle.

[0051] Optionally, the vehicle trajectory prediction device further includes a temporal convolutional network and a fusion module.

[0052] The temporal convolutional network is used to perform feature encoding on the historical motion information of the target vehicle to obtain the temporal features of the target vehicle; and to perform feature encoding on the historical motion information of the surrounding vehicles to obtain the temporal features of the surrounding vehicles;

[0053] The fully connected neural network uses one of several fully connected layers to encode the time series features of the target vehicle to obtain the query feature Q0;

[0054] The other two layers in the fully connected layers are used to encode the temporal features of the surrounding vehicles to obtain the Key feature K0 and the Value feature V0;

[0055] The fusion module fuses the query feature Q0, the key feature K0, and the value feature V0 according to the following formula to obtain the interaction feature:

[0056]

[0057] Among them, f interaction is the interaction feature, d o is the Key feature K o The feature dimension of ; T represents the matrix transpose.

[0058] Optionally, the vehicle trajectory prediction device further includes a PointNet network,

[0059] The PointNet network is used to perform feature encoding on lane centerline information to obtain lane line features;

[0060] The fully connected neural network uses one of several fully connected layers to encode the time series features of the target vehicle to obtain the query feature Q1;

[0061] The other two layers in the fully connected layers are used to encode the lane features respectively to obtain the Key feature K1 and the Value feature V1;

[0062] The fusion module fuses the query feature Q1, the key feature K1, and the value feature V1 according to the following formula to obtain the scene feature:

[0063]

[0064] Among them, f scene is the scene feature; d l is the Key feature K l The feature dimension of ; T represents the matrix transpose.

[0065] Optionally, the splicing module splices the interaction features, scene features, and time series features of the target vehicle according to the following formula:

[0066] f concat =[f target ;f interaction ;f scene ]

[0067] Among them, fconcat For fusion features.

[0068] Optionally, calculating future interaction features based on fused features includes the following steps:

[0069] Several fully connected layers are used to encode the fused features and obtain the initial predicted future trajectory;

[0070] The temporal convolutional network is used to extract the temporal features of the preliminary predicted future trajectory to obtain the temporal features of the preliminary predicted future trajectory;

[0071] Use one of several fully connected layers to encode the fusion features and obtain the query feature Q f ;

[0072] The other two layers in the fully connected layers are used to encode the temporal features of the predicted preliminary future trajectory to obtain the Key feature K f and Value feature V f ;

[0073] According to the following formula, the Query feature Q f , Key feature K f 、Value feature V f Fusion is performed to obtain future interaction features:

[0074]

[0075] Among them, f future_interact For future interactive features, d f is the Key feature K f The feature dimension of ; T represents the matrix transpose.

[0076] Optionally, the splicing module splices the fusion feature and the future interaction feature to form a coding feature;

[0077] The fully connected neural network uses several fully connected layers to decode the encoded features to form a predicted trajectory of the target vehicle.

[0078] Optionally, the splicing module splices the fusion feature and the future interaction feature according to the following formula:

[0079] f encoder =[f concat ;f future_interact ]

[0080] Among them, f encoder is the coding feature.

[0081] The beneficial effects of the vehicle trajectory prediction device provided by the present invention are similar to the beneficial effects reasoning process of the aforementioned vehicle trajectory prediction method, and will not be repeated here.

[0082] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements any of the above methods when executing the computer program.

[0083] At the same time, the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, any of the above methods is implemented.

[0084] Furthermore, the present invention further provides a motor vehicle having an automatic driving function, wherein when the automatic driving function is in operation, the motor vehicle predicts the vehicles around it by using any one of the vehicle trajectory prediction methods described above;

[0085] or the motor vehicle has any one of the aforementioned vehicle trajectory prediction devices;

[0086] or the motor vehicle has the aforementioned computer device;

[0087] Or the motor vehicle has the aforementioned computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by the processor, any one of the aforementioned vehicle trajectory prediction methods is implemented.

[0088] These features and advantages of the present invention will be further disclosed in the following detailed description and accompanying drawings. The preferred embodiments and means of the present invention will be fully illustrated in conjunction with the accompanying drawings, but are not intended to limit the technical solutions of the present invention. Furthermore, although multiple features, elements, and components may be present in each of the following text and accompanying drawings, they may be labeled with different symbols or numbers for convenience, but all represent components with the same or similar structure or function. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The present invention will be further described below in conjunction with the accompanying drawings:

[0090] Figure 1 This is a flow chart of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0091] The following is an explanation and description of the technical solutions of the embodiments of the present invention in conjunction with the drawings of the embodiments of the present invention. However, the following embodiments are only preferred embodiments of the present invention and are not exhaustive. Based on the embodiments in the implementation manner, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0092] Reference in this specification to "one embodiment," "an example," or "an example" means that a particular feature, structure, or characteristic described in connection with the embodiment itself can be included in at least one embodiment of the present patent disclosure. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0093] Example:

[0094] like Figure 1 As shown, this embodiment provides a vehicle trajectory prediction method for predicting the vehicle trajectory during automatic driving, including the following steps:

[0095] Interaction feature extraction: The interaction features between the target vehicle and its surrounding vehicles are obtained based on their historical motion information.

[0096] In this step, first, the historical motion information of the target vehicle and the vehicles around it is obtained respectively. The target vehicle is not a vehicle running autonomous driving, but refers to a vehicle that needs to be predicted around the vehicle running autonomous driving. The vehicle running autonomous driving uses the vehicle trajectory prediction method provided in this embodiment to predict the trajectory of each vehicle around it, and plans its own driving route according to the predicted trajectory to avoid scratches with surrounding vehicles. The historical motion information of the target vehicle and the vehicles around it can be obtained by the vehicle running autonomous driving through vision, laser radar, ultrasonic radar, etc. This is a prior art and will not be repeated here. In this embodiment, the historical motion information includes but is not limited to position coordinates, speed, acceleration, heading angle, etc. Other physical quantities that can be predicted by those skilled in the art to reflect the movement of the vehicle also belong to the historical motion information described in this embodiment and are not limited here.

[0097] The surrounding vehicles of the target vehicle mentioned in this embodiment refer to the vehicles that may be affected if the target vehicle needs to change lanes. For example, on a road with a fast speed and a small traffic flow, when a vehicle changes lanes to the left or right, it is necessary to consider the vehicle behind the left or right. In this case, the surrounding vehicles refer to the vehicle behind the left and the vehicle behind the right. On a road with a slow speed and a large traffic flow, when a vehicle changes lanes to the left or right, it is necessary not only to consider the vehicle behind the left or right, but also to consider the vehicles in the left front, right front, left side, and right side. In this case, the surrounding vehicles refer to the vehicles in the left front, left side, left rear, right front, right side, and right rear. Depending on the situation, it is also necessary to consider the vehicles directly in front and directly behind. For the selection of surrounding vehicles, similar to the currently mature lane change assistance system that has been widely used in the market, this is a prior art and will not be described in detail here.

[0098] A temporal convolutional network is used to encode the target vehicle's historical motion information to obtain its temporal features. The same network is used to encode the historical motion information of surrounding vehicles to obtain their temporal features. This method of obtaining vehicle temporal features using a temporal convolutional network is well known in the art and will not be further elaborated here.

[0099] One of the several fully connected layers is used to encode the time series features of the target vehicle to obtain the Query feature Q0; the other two layers of the several fully connected layers are used to encode the time series features of the surrounding vehicles to obtain the Key feature K0 and the Value feature V0. In this embodiment, the fully connected layer has three layers, wherein the Query feature Q0, the Key feature K0, and the Value feature V0 are obtained by different fully connected layers respectively. In other embodiments, the number of fully connected layers can also be increased or decreased by those skilled in the art according to actual conditions. The Query feature Q0, the Key feature K0, and the Value feature V0 are the general technical meanings in this field and will not be repeated here.

[0100] Using the cross-channel attention mechanism, the query feature Q0, key feature K0, and value feature V0 are fused according to the following formula to obtain the interactive feature:

[0101]

[0102] Among them, f interaction is the interaction feature, d o is the Key feature K o The feature dimension of ; T represents the matrix transpose.

[0103] Scene feature extraction: The scene features of the target vehicle and its adjacent lanes are obtained based on their historical motion information.

[0104] In this step, the lane centerline information of the target vehicle's adjacent lane is obtained. The target vehicle's adjacent lane refers to the lane on the left or right side where the target vehicle can directly merge into.

[0105] The lane centerline information is feature-encoded using the PointNet network to obtain lane features. The PointNet network is well-known in the art, and obtaining lane features using the PointNet network is also well-known in the art and will not be further described here.

[0106] One of the several fully connected layers is used to encode the temporal features of the target vehicle to obtain the Query feature Q1, and the other two layers of the several fully connected layers are used to encode the lane line features respectively to obtain the Key feature K1 and the Value feature V1. In this embodiment, since the parameter values ​​and weight values ​​in the fully connected layer have changed after different steps, the Query feature Q0 and the Query feature Q1 calculated using the fully connected layer are different. Similarly, the Key feature K1 and the Value feature V1 obtained by the fully connected layer are also different from the Key feature K0 and the Value feature V0. This is a prior art known in the art. At the same time, the Query feature Q1, the Key feature K1 and the Value feature V1 are also the general technical meanings in this field and will not be repeated here.

[0107] Using the cross-channel attention mechanism, the query feature Q1, key feature K1, and value feature V1 are fused according to the following formula to obtain the scene feature:

[0108]

[0109] Among them, f scene is the scene feature; d l is the Key feature K l The feature dimension of ; T represents the matrix transpose.

[0110] Feature splicing: Splice the interaction features, scene features, and target vehicle’s temporal features to form a fusion feature. In this step, the interaction features, scene features, and target vehicle’s temporal features are spliced ​​according to the following formula:

[0111] f concat =[f target ;f interaction ;f scene ]

[0112] Among them, f concat For fusion features.

[0113] After the fusion features are formed, the future interaction features are calculated based on the fusion features. In this step, the fusion features are encoded using several fully connected layers. The fusion features are calculated in turn through each fully connected layer. The output of the previous fully connected layer is used as the input of the next fully connected layer to obtain a preliminary prediction of the future trajectory. The temporal features of the preliminary predicted future trajectory are extracted using a temporal convolutional network to obtain the temporal features of the preliminary predicted future trajectory. Those skilled in the art will know that using a temporal network to obtain the temporal features of the preliminary predicted future trajectory is also a prior art.

[0114] Use one of several fully connected layers to encode the fusion features and obtain the query feature Q f , use the other two layers in several fully connected layers to encode the temporal features of the predicted preliminary future trajectory and obtain the Key feature K f and Value feature V f . Get the Query feature Q through the fully connected layer f , Key feature K f and Value feature V f It is a well-known prior art in the art. f , Key feature K f and Value feature V f This is also the common technical meaning in this field and will not be repeated here.

[0115] According to the following formula, the Query feature Q f , Key feature K f 、Value feature V f Fusion is performed to obtain future interaction features:

[0116]

[0117] Among them, f future_interact For future interactive features, d f is the Key feature K f The feature dimension of ; T represents the matrix transpose.

[0118] Finally, the fused features and future interaction features are concatenated and decoded to form the predicted trajectory of the target vehicle. In this step, the fused features and future interaction features are concatenated to form the encoded features. The concatenation method is the same as the concatenation method for the interaction features, scene features, and target vehicle temporal features mentioned above, using the following formula:

[0119] f encoder =[f concat ;f future_interact ]

[0120] Among them, f encoder is the coding feature.

[0121] The encoded features are then decoded using several fully connected layers to form the predicted trajectory of the target vehicle.

[0122] The vehicle trajectory prediction method provided in this embodiment combines a temporal convolutional neural network with a cross-channel attention mechanism, taking into account the interaction information between vehicles and the environmental information between vehicles and lanes, accurately extracting the interaction features of surrounding vehicles and the scene features around the vehicle, making the trajectory prediction results of surrounding vehicles more accurate, thereby ensuring that the autonomous vehicle completes reasonable trajectory planning and controls the vehicle to drive smoothly.

[0123] At the same time, this embodiment also provides a vehicle trajectory prediction device, which is used to execute the aforementioned vehicle trajectory prediction method during automatic vehicle driving and predict the vehicle trajectory, including: a temporal convolutional network, a PointNet network, a fully connected neural network, a temporal convolutional network, a fusion module and a splicing module.

[0124] The temporal convolutional network is used to feature encode the historical motion information of the target vehicle to obtain the temporal features of the target vehicle; feature encode the historical motion information of the surrounding vehicles to obtain the temporal features of the surrounding vehicles; and perform temporal feature extraction on the preliminary predicted future trajectory to obtain the temporal features of the preliminary predicted future trajectory.

[0125] The PointNet network is used to encode the lane centerline information and obtain lane line features.

[0126] A fully connected neural network is used to derive interactive features between a target vehicle and its surrounding vehicles based on their historical motion information. Specifically, the neural network uses one of several fully connected layers to encode the target vehicle's temporal features, resulting in the query feature Q0. Two other fully connected layers encode the temporal features of the surrounding vehicles, resulting in the key feature K0 and the value feature V0. The specific number of fully connected layers can be flexibly selected by those skilled in the art based on practical needs. This is conventional technology and is not intended to be limiting here.

[0127] The fully connected neural network is also used to derive scene features from the target vehicle's historical motion information and the centerline information of its adjacent lanes. Specifically, the neural network uses one of several fully connected layers to encode the target vehicle's temporal features, generating the query feature Q1. Two other fully connected layers encode the lane line features, generating the key feature K1 and the value feature V1.

[0128] The fully connected neural network is also used to encode the fusion features using one of several fully connected layers to obtain the query feature Q f ; Use the other two layers in several fully connected layers to encode the temporal features of the predicted preliminary future trajectory to obtain the Key feature K f and Value feature Vf .

[0129] Since the parameter values ​​and weight values ​​in the fully connected layer have changed after different steps, the Query feature Q0, Query feature Q1 and Query feature Q calculated by the fully connected layer are f Different, similarly, Key feature K o , Key feature K1 and Key feature K f There are also differences in Value feature V0, Value feature V1 and Value feature V f This is also different, and this is prior art, so I will not go into details here.

[0130] The fusion module is used to fuse the query feature Q0, key feature K0, and value feature V0 according to the following formula to obtain the interaction feature:

[0131]

[0132] Among them, f interaction is the interaction feature, d o is the Key feature K o The characteristic dimension of ; T represents the matrix transpose;

[0133] The query feature Q1, key feature K1, and value feature V1 are fused according to the following formula to obtain the scene feature:

[0134]

[0135] Among them, f scene is the scene feature; d l is the Key feature K l The characteristic dimension of ; T represents the matrix transpose;

[0136] And according to the following formula, the Query feature Q f , Key feature K f 、Value feature V f Fusion is performed to obtain future interaction features:

[0137]

[0138] Among them, f future_interact For future interactive features, d f is the Key feature K f The feature dimension of ; T represents the matrix transpose.

[0139] The splicing module is used to splice the interaction features, scene features, and target vehicle temporal features according to the following formula to form a fusion feature:

[0140] f concat =[f target ;f interaction ;f scene ]

[0141] Among them, f concat For fusion features.

[0142] And the fusion features and future interaction features are spliced ​​according to the following formula to form the encoding features:

[0143] f encoder =[f concat ;f future_interact ]

[0144] Among them, f encoder is the coding feature.

[0145] The fully connected neural network is also used to calculate future interaction features based on the fused features, and to concatenate and decode the fused features and future interaction features into the predicted trajectory of the target vehicle.

[0146] This embodiment also provides a computer device including a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the vehicle trajectory prediction method described above. The steps of the vehicle trajectory prediction method can be the steps of the memory analysis method described in each of the above embodiments.

[0147] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program. Accordingly, the computer program can be stored in a non-volatile computer-readable storage medium, and when the computer program is executed, it can implement the method of any of the above-mentioned embodiments. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0148] Furthermore, this embodiment also provides a motor vehicle having an automatic driving function. When the motor vehicle provided by this embodiment operates the automatic driving function, the vehicle trajectory prediction method described above predicts the vehicles around it;

[0149] Or the motor vehicle provided in this embodiment has the aforementioned vehicle trajectory prediction device;

[0150] Or the motor vehicle provided in this embodiment has the aforementioned computer device;

[0151] Or the motor vehicle provided in this embodiment has the aforementioned computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the aforementioned vehicle trajectory prediction method is implemented.

[0152] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes but is not limited to the contents described in the drawings and the above specific embodiments. Any modifications that do not deviate from the functional and structural principles of the present invention are intended to be included within the scope of the claims.

Claims

1. A vehicle trajectory prediction method for predicting a vehicle trajectory during automatic driving, characterized in that: The vehicle trajectory prediction method comprises the following steps: Obtain the interaction features between the target vehicle and its surrounding vehicles based on their historical motion information; The scene features of both the target vehicle and its adjacent lanes are obtained based on their historical motion information. Combine the interaction features, scene features, and temporal features of the target vehicle to form fusion features; Calculate future interaction features based on fused features; The fused features and future interaction features are concatenated and decoded to form the predicted trajectory of the target vehicle; The calculating of future interaction features based on the fusion features includes: Several fully connected layers are used to encode the fused features and obtain the initial predicted future trajectory; The temporal convolutional network is used to extract the temporal features of the preliminary predicted future trajectory to obtain the temporal features of the preliminary predicted future trajectory; Use one of several fully connected layers to encode the fusion features and obtain the query feature Q f ; The other two layers in the fully connected layers are used to encode the temporal features of the predicted preliminary future trajectory to obtain the Key feature K f and Value feature V f ; According to the following formula, the Query feature Q f , Key feature K f 、Value feature V f Fusion is performed to obtain future interaction features: Among them, f future_interact For future interactive features, d f is the Key feature K f The feature dimension of ; T represents the matrix transpose.

2. The vehicle trajectory prediction method according to claim 1, characterized in that: The following steps are used to obtain the interaction features between the target vehicle and its surrounding vehicles based on their historical motion information: Obtain the historical motion information of the target vehicle and its surrounding vehicles respectively; The temporal convolutional network is used to encode the historical motion information of the target vehicle and obtain the temporal characteristics of the target vehicle; Use the temporal convolutional network to encode the historical motion information of surrounding vehicles and obtain the temporal features of surrounding vehicles; Use one of several fully connected layers to encode the temporal features of the target vehicle to obtain the query feature Q0; The other two layers in the fully connected layers are used to encode the temporal features of the surrounding vehicles to obtain the Key feature K0 and the Value feature V0; The query feature Q0, key feature K0, and value feature V0 are fused according to the following formula to obtain the interactive feature: Among them, f interaction is the interaction feature, d o is the Key feature K o characteristic dimensions; T represents matrix transpose.

3. The vehicle trajectory prediction method according to claim 2, characterized in that: The scene features of the target vehicle and its adjacent lanes are obtained based on their historical motion information, including the following steps: Obtain lane centerline information of adjacent lanes of the target vehicle; Use the PointNet network to encode the lane centerline information and obtain lane line features; Use one of several fully connected layers to encode the temporal features of the target vehicle to obtain the query feature Q1; The other two layers in the fully connected layers are used to encode the lane features respectively to obtain the Key feature K1 and the Value feature V1; The query feature Q1, key feature K1, and value feature V1 are fused according to the following formula to obtain the scene feature: Among them, f scene is the scene feature; d l is the Key feature K l The feature dimension of ; T represents the matrix transpose.

4. The vehicle trajectory prediction method according to claim 3, characterized in that: The interaction features, scene features, and target vehicle temporal features are concatenated according to the following formula: f concat =[f target ;f interaction ;f scene ] Among them, f concat For fusion features.

5. The vehicle trajectory prediction method according to claim 4, characterized in that: The steps of concatenating and decoding the fused features and future interaction features to form the predicted trajectory of the target vehicle include the following: Concatenate the fusion features and future interaction features to form encoding features; Several fully connected layers are used to decode the encoded features to form the predicted trajectory of the target vehicle.

6. The vehicle trajectory prediction method according to claim 1, characterized in that: The fusion features and future interaction features are spliced ​​according to the following formula: f encoder =[f concat ;f future_interact ] Among them, f encoder is the coding feature.

7. A vehicle trajectory prediction device for predicting the vehicle trajectory during automatic driving, characterized in that: The vehicle trajectory prediction device comprises: A fully connected neural network is used to obtain interaction features between the target vehicle and its surrounding vehicles based on their historical motion information; and to obtain scene features of the target vehicle and its adjacent lanes based on their historical motion information and lane centerline information. A splicing module is used to splice the interaction features, scene features and the time series features of the target vehicle to form a fusion feature; and to splice the fusion feature and the future interaction features; The fully connected neural network is further used to calculate future interaction features based on the fused features; and to decode the fused features and the future interaction features after splicing them together to form a predicted trajectory of the target vehicle; Calculating future interaction features based on fusion features includes the following steps: Several fully connected layers are used to encode the fused features and obtain the initial predicted future trajectory; The temporal convolutional network is used to extract the temporal features of the preliminary predicted future trajectory to obtain the temporal features of the preliminary predicted future trajectory; Use one of several fully connected layers to encode the fusion features and obtain the query feature Q f The other two layers in the fully connected layers are used to encode the temporal features of the predicted preliminary future trajectory to obtain the Key feature K f and Value feature V f ; According to the following formula, the Query feature Q f , Key feature K f 、Value feature V f Fusion is performed to obtain future interaction features: Among them, f future_interact For future interactive features, d f is the Key feature K f The feature dimension of ; T represents the matrix transpose.

8. The vehicle trajectory prediction device according to claim 7, characterized in that: The vehicle trajectory prediction device also includes a temporal convolutional network and a fusion module. The temporal convolutional network is used to perform feature encoding on the historical motion information of the target vehicle to obtain the temporal features of the target vehicle; Perform feature encoding on the historical motion information of surrounding vehicles to obtain the temporal characteristics of surrounding vehicles; The fully connected neural network uses one of several fully connected layers to encode the time series features of the target vehicle to obtain the query feature Q0; The other two layers in the fully connected layers are used to encode the temporal features of the surrounding vehicles to obtain the Key feature K0 and the Value feature V0; The fusion module fuses the query feature Q0, the key feature K0, and the value feature V0 according to the following formula to obtain the interaction feature: Among them, f interaction is the interaction feature, d o is the Key feature K o characteristic dimensions; T represents matrix transpose.

9. The vehicle trajectory prediction device according to claim 8, characterized in that: The vehicle trajectory prediction device also includes a PointNet network, The PointNet network is used to perform feature encoding on lane centerline information to obtain lane line features; The fully connected neural network uses one of several fully connected layers to encode the time series features of the target vehicle to obtain the query feature Q1; The other two layers in the fully connected layers are used to encode the lane features respectively to obtain the Key feature K1 and the Value feature V1; The fusion module fuses the query feature Q1, the key feature K1, and the value feature V1 according to the following formula to obtain the scene feature: Among them, f scene is the scene feature; d l is the Key feature K l The feature dimension of ; T represents the matrix transpose.

10. The vehicle trajectory prediction device according to claim 9, characterized in that: The splicing module splices the interaction features, scene features and target vehicle temporal features according to the following formula: f concat =[f target ;f interaction ;f scene ] Among them, f concat For fusion features.

11. The vehicle trajectory prediction device according to claim 7, characterized in that: The splicing module splices the fusion features and the future interaction features to form coding features; The fully connected neural network uses several fully connected layers to decode the encoded features to form a predicted trajectory of the target vehicle.

12. The vehicle trajectory prediction device according to claim 11, characterized in that: The splicing module splices the fusion features and the future interaction features according to the following formula: f encoder =[f concat ;f future_interact ] Among them, f encoder is the coding feature.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the vehicle trajectory prediction method according to any one of claims 1 to 6 is implemented.

14. 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 trajectory prediction method according to any one of claims 1 to 6 is implemented.

15. A motor vehicle, characterized in that: The motor vehicle has an automatic driving function, and when the automatic driving function is running, the motor vehicle predicts the vehicles around it by using the vehicle trajectory prediction method according to any one of claims 1 to 6; or the motor vehicle has the vehicle trajectory prediction device according to any one of claims 7 to 12; or the motor vehicle has the computer device according to claim 13; Or the motor vehicle has the computer-readable storage medium according to claim 14, and when the computer program is executed by a processor, the vehicle trajectory prediction method according to any one of claims 1 to 6 is implemented.

Images