Knowledge-fused multi-stage target guidance vehicle trajectory prediction method

Abstract

The invention discloses a knowledge-fused multi-stage target guidance vehicle trajectory prediction method. The method comprises the following steps: preprocessing vehicle information and a local high-precision map; illegal lane segments are filtered based on traffic rules and domain knowledge; vehicle and map information coding and society interaction capture; the invention relates to multi-stage target prediction and target guidance-based decoding. According to the invention, based on the fact that a plurality of intermediate targets can be planned in advance when a vehicle goes to a destination, multi-stage target lane segments are estimated along the way to guide and restrain trajectory prediction. By using reasonable prior and beneficial information provided by a target lane section, accumulative errors in a prediction process are reduced, so that scene consistency and stability of trajectory prediction are improved. According to the method, a dynamic target lane section screening area is designed by integrating domain knowledge and traffic rules, interference of a large number of illegal false candidate targets is filtered, and faster and more accurate trajectory prediction is achieved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving trajectory prediction, and in particular to a vehicle trajectory prediction method considering fusion knowledge and multi-stage target guidance. Background Art

[0002] Accurately predicting the trajectories of surrounding vehicles is essential for autonomous driving systems to achieve efficient, safe, and anthropomorphic motion planning. In interactive environments, the key challenges of trajectory prediction are to capture the uncertainty of moving targets and ensure the scene consistency of the prediction results. In recent years, the main steps of goal-guided vehicle trajectory prediction methods are: first determine the vehicle's moving targets, and then predict the trajectory based on these targets. This method significantly improves the performance of long-term trajectory prediction by associating the future trajectory distribution with the vehicle's target intention. However, most current goal-guided vehicle trajectory prediction methods only estimate a final long-term target point, ignoring that the vehicle's moving target will change dynamically over time and that multiple intermediate targets are usually planned in advance during driving. In addition, the vehicle will select potential targets that conform to the road topology and traffic rules based on the current state and surrounding environment, rather than indiscriminately paying attention to all physical spaces. However, existing methods usually sample target points in an unconstrained two-dimensional space, ignoring the fact that vehicle motion is constrained by road geometry, map topology, and traffic rules. The above two problems usually cause the predicted trajectory to deviate from the lane or violate traffic rules, thereby reducing the prediction scene consistency and stability. Therefore, how to integrate domain knowledge and traffic rules, design a dynamic target screening area to filter out a large number of illegal false candidate targets, and estimate multiple reasonable intermediate targets in this area to guide and constrain vehicle trajectory prediction, thereby promoting the autonomous driving system to make stable and reliable driving decisions, is an urgent problem to be solved. Summary of the invention

[0003] In order to solve the above problems existing in the prior art, the present invention proposes a multi-stage target-guided vehicle trajectory prediction method that integrates knowledge, utilizes multi-stage target lane segment constraints and guided vehicle trajectory prediction, and designs a dynamic target screening area by integrating domain knowledge and traffic rules to filter out a large number of illegal false candidate targets, thereby achieving faster and more accurate trajectory prediction while improving the scene consistency and stability of the prediction.

[0004] In order to achieve the above object, the technical solution of the present invention is as follows: a multi-stage target-guided vehicle trajectory prediction method integrating knowledge, comprising the following steps:

[0005] A. Vehicle information and local high-precision map preprocessing

[0006] A1. Use the perception and positioning technology of the autonomous driving system to obtain the historical data T of N vehicles around hInformation at that moment as well as High-precision map information within meters The surrounding N vehicles include the vehicle itself. All information is normalized into a local coordinate system, the origin of the local coordinate system is set as the current position of the target vehicle, and the longitudinal forward direction of the target vehicle is defined as the positive direction of the y-axis of the local coordinate system.

[0007] A2. For the surrounding vehicle i, connect its adjacent time information into a vector. Therefore, the history T of the surrounding vehicle i h The information at each moment is represented as T h -1 vector collection s i , the vector dimension is set to V dim ,Right now:

[0008]

[0009] In the formula, Superscript t∈[-T h +2,0] is the timestamp, and the subscript i∈[1,N] is the number of the surrounding vehicles. is the starting position of the vector, is the x-axis coordinate of the starting point of the vector, is the y-axis coordinate of the starting point of the vector, is the vector end position, is the x-axis coordinate of the vector's end point, is the y-coordinate of the vector’s end point, It's speed. It is the direction, yes

[0010] acceleration, onehot(i) is the unique hot encoding for the vehicle, target vehicle, and other vehicles, pad veh is zero-filled. Finally, N surrounding vehicles T h The information at each moment is represented as follows:

[0011]

[0012] In the formula, Represents the vector features of the surrounding vehicle i at time t.

[0013] A3. For local high-precision maps, all lane centerlines are divided into L non-overlapping segments, each consisting of D equidistant points. For lane segment l, adjacent points are connected into vectors. Therefore, D equidistant points are represented as a set of D-1 vectors m l , the vector dimension is set to V dim , which is expressed as follows:

[0014]

[0015] In the formula, The superscript d∈[1,D-1] is the number of the equidistant point, and the subscript l∈[1,L] is the lane segment number. is the starting position of the vector, is the x-axis coordinate of the starting point of the vector, is the y-axis coordinate of the starting point of the vector, is the vector end position, is the x-axis coordinate of the vector's end point, is the y-coordinate of the vector’s end point, is the vector yaw angle, is a turning direction, wherein the turning direction includes left turn, right turn and straight ahead, Indicates whether it is at an intersection. Indicates whether there is a back neighbor, Indicates whether there is a front neighbor, idx is the index of the lane segment, pad map is zero-filled. Finally, L lane segments consisting of D equidistant points are represented as follows:

[0016]

[0017] In the formula, The vector feature representing the dth equidistant point in lane segment l.

[0018] B. Filter out illegal lane segments based on traffic rules and domain knowledge

[0019] B1. Determine the seed lane segment based on traffic rules

[0020] Calculate the positions of equally spaced points for each lane segment The current position of the target vehicle The minimum distance Δd, subscript l is the number of the lane segment, superscript d is the number of the equidistant point, subscript tar represents the target vehicle, and superscript 0 represents the current moment. Then, the yaw angle of each equidistant point in the lane segment is calculated. The current direction of the target vehicle The minimum difference Δy is expressed as follows:

[0021]

[0022]

[0023] Where dist(·) is the function for calculating the Euclidean distance, and min(·) is the function for taking the minimum value.

[0024] In the first stage of determining the seed lane segment, the condition Δd < τ will be satisfied d And Δy<τy The lane segment of is marked as a candidate seed lane segment, where τ d is the set distance threshold, τ y In the second stage of determining the seed lane segment, according to the traffic rules and lane topology, the candidate seed lane segment closest to the current position of the target vehicle in each lane is regarded as the seed lane segment Seed. lane , the seed lane segment comes from the current lane or the adjacent lane.

[0025] B2. Searching for legal lane segments based on domain knowledge

[0026] First, according to the current state of the target vehicle and kinematic knowledge, the forward search depth τ is determined succ and the backward search depth τ prec , which is expressed as follows:

[0027]

[0028] τ prec =δ

[0029] Where, T f is the predicted duration, The current speed of the target vehicle, The current acceleration of the target vehicle, 0<δ<500 is a fixed constant.

[0030] Then, the seed lane segment Seed determined in step B1 is used lane As the search starting point, obtain the legal lane segment based on the breadth-first search algorithm It is expressed as follows:

[0031]

[0032]

[0033]

[0034] In the formula, BFS(·) is the breadth-first search algorithm, Link succ is the forward adjacency relationship of the lane, Link prec is the lane rearward adjacency relationship, It is the forward legal lane segment, It is a rearward legal lane section. is all legal lane segments. Finally, step A3 preprocessing is performed

[0035] C. Vehicle and map information encoding and social interaction capture

[0036] C1. Based on the unified representation of vehicle history information and high-precision map information, three encoders with the same structure and different weights, which are composed of gated multilayer perceptrons (gMLP), are used to encode the vector sets represented by steps A2, A3 and B2 respectively. and Extract vehicle features f v 、All lane segment features f m and legal lane segment characteristics It is expressed as follows:

[0037]

[0038]

[0039]

[0040] Where VehicleNet(·) is the vehicle feature encoder, W v is the weight for extracting vehicle features, MapNet(·) is the map feature encoder, and W m is the weight for extracting features of all lane segments, is the weight for extracting the features of legal lane segments.

[0041] C2, vehicle feature f extracted based on step C1 v 、All lane segment features f m and legal lane segment characteristics The vehicle-vehicle interactor composed of self-attention mechanism is used to capture the vehicle-vehicle interaction features f v2v , a vehicle-map interactor consisting of a cross-attention mechanism is used to capture the vehicle-map interaction features f v2m , a map-map interactor consisting of a cross-attention mechanism is used to capture map-map interaction features It is expressed as follows:

[0042] f v2v =SA(f v ,f v )

[0043] f v2m =CA(f v2v ,f m )

[0044]

[0045] Where, SA(·) is the self-attention mechanism module, and CA(·) is the cross-attention mechanism module.

[0046] D. Multi-stage target prediction and target-guided decoding

[0047] D1, based on the fact that the vehicle plans multiple intermediate goals in advance while moving to ensure safe arrival at the destination, the legal lane segment features are used Vehicle-map interaction features v2m and map-map interaction features Estimate the target lane segment of group K in stage J from the legal lane segment It is expressed as follows:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] In the formula, φ cat (·) is the feature concatenation operation, is a multi-layer perceptron, or MLP, softmax(·) is the softmax function, and score j is the probability score table of the jth stage, argtop(·,k) is the index function for taking the top k values, and index(·) is the function for taking values ​​by index.

[0054] D2 follows the paradigm of the target guidance method and utilizes the multi-stage target lane segment guidance and constrained trajectory prediction of step D1.

[0055] First, to promote the consistency and stability of the prediction scene, the multi-stage target lane segments are integrated through the cross-attention mechanism. and vehicle-map interaction v2m Fusion is performed to obtain fusion features It is expressed as follows:

[0056]

[0057] Then, in order to prevent the predicted trajectory and the target lane segment from deviating from each other over time, a dynamic adaptive adjuster is constructed based on the learnable query, which is expressed as follows:

[0058]

[0059] Where Φ is the learnable query, ⊙ is the element-wise dot product operation, and Θ is the output feature of the dynamic adaptive adjuster.

[0060] Finally, to enhance the deep understanding of scene information, the map-map interaction feature As the initial hidden state of the gated recurrent unit, GRU, and as one of the conditions for probability prediction. Multiple trajectories are obtained by GRU decoding, and the related probabilities are obtained by MLP decoding, which are expressed as follows:

[0061]

[0062]

[0063] In the formula, Υ k is the predicted k-th trajectory, p k is the probability of the predicted kth track, and k is the number of the predicted track.

[0064] Furthermore, the breadth-first search algorithm in step B2 is replaced by a depth-first search algorithm or a Monte Carlo tree search algorithm.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] 1. Based on the fact that the vehicle will plan multiple intermediate goals in advance on the way to the destination, the present invention estimates multiple stages of target lane segments along the way to guide and constrain trajectory prediction. By using the reasonable prior and useful information provided by the target lane segments, the cumulative error in the prediction process is reduced, thereby improving the scene consistency and stability of trajectory prediction.

[0067] 2. The present invention takes into account that vehicles usually select targets that conform to the road topology and traffic rules based on the current state and surrounding environment. By integrating domain knowledge and traffic rules to design a dynamic target lane segment screening area, the interference of a large number of illegal false candidate targets is filtered out to achieve faster and more accurate trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the process of the present invention.

[0069] Figure 2 It is a schematic diagram of the multi-stage target lane segment guidance and constraint trajectory prediction process.

[0070] Figure 3 It is a structural diagram of a multi-stage target-guided vehicle trajectory prediction model that integrates knowledge.

[0071] Figure 4 This is the structure diagram of the vehicle feature encoder and map feature encoder.

[0072] Figure 5 It is a structural diagram of the multi-stage target lane segment predictor.

[0073] Figure 6 It is the structural diagram of the target-guided decoder. DETAILED DESCRIPTION

[0074] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings. Figure 1 As shown, a multi-stage target-guided vehicle trajectory prediction method integrating knowledge includes the following steps:

[0075] A. Vehicle information and local high-precision map preprocessing

[0076] A1. Use the perception and positioning technology of the autonomous driving system to obtain the historical data T of N vehicles around h Information at that moment As well as high-precision map information within the range of 80 meters in front, 20 meters behind, 50 meters on the left, and 50 meters on the right The surrounding N vehicles include the vehicle itself. All information is normalized to a local coordinate system, the origin of which is set as the current position of the target vehicle, and the longitudinal direction of the target vehicle is defined as the positive direction of the y-axis of the local coordinate system. h =20, the interval between adjacent moments is 0.1 seconds.

[0077] A2. For the surrounding vehicle i, connect its adjacent time information into a vector. Therefore, the history T of the surrounding vehicle i h The information at each moment is represented as T h -1 vector collection s i , the vector dimension is set to V dim ,Right now:

[0078]

[0079] In the formula, Superscript t∈[-T h +2,0] is the timestamp, and the subscript i∈[1,N] is the number of the surrounding vehicles. is the starting position of the vector, is the x-axis coordinate of the starting point of the vector, is the y-axis coordinate of the starting point of the vector, is the vector end position, is the x-axis coordinate of the vector's end point, is the y-coordinate of the vector’s end point, It's speed. It is the direction, is the acceleration, onehot(i) is the unique hot encoding for the vehicle, target vehicle and other vehicles, pad veh is zero-filled. Finally, N surrounding vehicles T h The information at each moment is represented as follows:

[0080]

[0081] In the formula, represents the vector features of the surrounding vehicle i at time t. In the embodiment of the present invention, V dim =16.

[0082] A3. For local high-precision maps, all lane centerlines are divided into L non-overlapping segments, each consisting of D equidistant points. For lane segment l, adjacent points are connected into vectors. Therefore, D equidistant points are represented as a set of D-1 vectors m l , the vector dimension is set to V dim , which is expressed as follows:

[0083]

[0084] In the formula, The superscript d∈[1,D-1] is the number of the equidistant point, and the subscript l∈[1,L] is the lane segment number. is the starting position of the vector, is the x-axis coordinate of the starting point of the vector, is the y-axis coordinate of the starting point of the vector, is the vector end position, is the x-axis coordinate of the vector's end point, is the y-coordinate of the vector’s end point, is the vector yaw angle, is a turning direction, wherein the turning direction includes left turn, right turn and straight ahead, Indicates whether it is at an intersection. Indicates whether there is a back neighbor, Indicates whether there is a front neighbor, idx is the index of the lane segment, pad map is zero-filled. Finally, L lane segments consisting of D equidistant points are represented as follows:

[0085]

[0086] In the formula, represents the vector feature of the dth equidistant point in lane segment l. In the embodiment of the present invention, D=5, and the resolution of the equidistant points is 1 meter.

[0087] B. Filter out illegal lane segments based on traffic rules and domain knowledge

[0088] B1. Determine the seed lane segment based on traffic rules

[0089] Calculate the positions of equally spaced points for each lane segment The current position of the target vehicle The minimum distance Δd, subscript l is the number of the lane segment, superscript d is the number of the equidistant point, subscript tar represents the target vehicle, and superscript 0 represents the current moment. Then, the yaw angle of each equidistant point in the lane segment is calculated. The current direction of the target vehicle The minimum difference Δy is expressed as follows:

[0090]

[0091]

[0092] Where dist(·) is the function for calculating the Euclidean distance, and min(·) is the function for taking the minimum value.

[0093] In the first stage of determining the seed lane segment, the condition Δd < τ will be satisfied d And Δy<τ y The lane segment of is marked as a candidate seed lane segment, where τ d =8 is the set distance threshold, τ y = 60° is the set orientation difference threshold. In the second stage of determining the seed lane segment, according to the traffic rules and lane topology, the candidate seed lane segment closest to the current position of the target vehicle in each lane is regarded as the seed lane segment Seed. lane , the seed lane segment comes from the current lane or the adjacent lane.

[0094] B2. Searching for legal lane segments based on domain knowledge

[0095] First, according to the current state of the target vehicle and kinematic knowledge, the forward search depth τ is determined succ and the backward search depth τ prec , which is expressed as follows:

[0096]

[0097] τ prec =δ

[0098] Where, T f is the predicted duration, The current speed of the target vehicle, The current acceleration of the target vehicle, δ=20 is a fixed constant. f = 3 seconds, and the interval between adjacent moments is 0.1 seconds.

[0099] Then, the seed lane segment Seed determined in step B1 is used lane As the search starting point, obtain the legal lane segment based on the breadth-first search algorithm It is expressed as follows:

[0100]

[0101]

[0102]

[0103] In the formula, BFS(·) is the breadth-first search algorithm, Link succ is the forward adjacency relationship of the lane, Link prec is the lane rearward adjacency relationship, It is the forward legal lane segment, It is a rearward legal lane section. All legal lane sections, all legal lane sections show opinions Figure 2 shown.

[0104] Finally, step A3 pretreatment is used

[0105] C. Vehicle and map information encoding and social interaction capture

[0106] C1. Based on the unified representation of vehicle history information and high-precision map information, three encoders with the same structure and different weights, which are composed of gated multilayer perceptrons (gMLP), are used to encode the vector sets represented by steps A2, A3 and B2 respectively. and Extract vehicle features f v 、All lane segment features f m and legal lane segment characteristics It is expressed as follows:

[0107]

[0108]

[0109]

[0110] Where VehicleNet(·) is the vehicle feature encoder, W v is the weight for extracting vehicle features, MapNet(·) is the map feature encoder, and W m is the weight for extracting features of all lane segments, is the weight for extracting the features of the legal lane segment. The detailed structure of the multi-stage target-guided vehicle trajectory prediction model integrating knowledge in the embodiment of the present invention is as follows: Figure 3 As shown, the detailed structures of the vehicle feature encoder and map feature encoder are as follows Figure 4 shown.

[0111] C2. See Figure 3 As shown, based on the vehicle feature f extracted in step C1 v、All lane segment features f m and legal lane segment characteristics The vehicle-vehicle interactor composed of self-attention mechanism is used to capture the vehicle-vehicle interaction features f v2v , a vehicle-map interactor consisting of a cross-attention mechanism is used to capture the vehicle-map interaction features f v2m , a map-map interactor consisting of a cross-attention mechanism is used to capture map-map interaction features It is expressed as follows:

[0112] f v2v =SA(f v ,f v )

[0113] f v2m =CA(f v2v ,f m )

[0114]

[0115] Where, SA(·) is the self-attention mechanism module, and CA(·) is the cross-attention mechanism module.

[0116] D. Multi-stage target prediction and target-guided decoding

[0117] D1. See Figure 5 As shown in the figure, based on the fact that the vehicle plans multiple intermediate goals in advance while moving to ensure safe arrival at the destination, the legal lane segment features are used Vehicle-map interaction features v2m and map-map interaction features Estimate the target lane segment of group K in stage J from the legal lane segment It is expressed as follows:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] In the formula, φ cat (·) is the feature concatenation operation, is a multi-layer perceptron, or MLP, softmax(·) is the softmax function, and score jis the probability score table of the jth stage, argtop(·,k) is the index function for taking the top k values, and index(·) is the function for taking values ​​according to the index. In the embodiment of the present invention, K=6, J=10.

[0124] D2 follows the paradigm of the target guidance method and utilizes the multi-stage target lane segment guidance and constrained trajectory prediction of step D1.

[0125] First, see Figure 5 As shown, in order to promote the consistency and stability of the predicted scene, the multi-stage target lane segments are integrated through the cross-attention mechanism. and vehicle-map interaction v2m Fusion is performed to obtain fusion features It is expressed as follows:

[0126]

[0127] Then, see Figure 5 As shown in the figure, in order to prevent the predicted trajectory and the target lane segment from deviating from each other over time, a dynamic adaptive adjuster is constructed based on the learnable Query, which is expressed as follows:

[0128]

[0129] Where Φ is the learnable query, ⊙ is the element-wise dot product operation, and Θ is the output feature of the dynamic adaptive adjuster.

[0130] Finally, see Figure 5 As shown in the figure, in order to enhance the deep understanding of scene information, the map-map interaction feature As the initial hidden state of the gated recurrent unit, GRU, and as one of the conditions for probability prediction. Multiple trajectories are obtained by GRU decoding, and the related probabilities are obtained by MLP decoding, which are expressed as follows:

[0131]

[0132]

[0133] In the formula, Υ k is the predicted k-th trajectory, p k is the probability of the predicted kth track, and k is the number of the predicted track.

[0134] Furthermore, the breadth-first search algorithm in step B2 may be replaced by a depth-first search algorithm or a Monte Carlo tree search algorithm.

[0135] The above describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected.

Claims

1. A multi-stage target-guided vehicle trajectory prediction method integrating knowledge, characterized by: The following steps are involved: A. Vehicle information and local high-precision map preprocessing A1. Use the perception and positioning technology of the autonomous driving system to obtain the historical data T of N vehicles around h Information at that moment as well as High-precision map information within meters The N surrounding vehicles include the vehicle itself; and all information is normalized into a local coordinate system, the origin of the local coordinate system is set as the current position of the target vehicle, and the longitudinal forward direction of the target vehicle is defined as the positive direction of the y-axis of the local coordinate system; A2. For the surrounding vehicle i, connect its adjacent time information into a vector; therefore, the history T of the surrounding vehicle i h The information at each moment is represented as T h -1 vector collection s i , the vector dimension is set to V dim ,Right now: In the formula, Superscript t∈[-T h +2,0] is the timestamp, and the subscript i∈[1,N] is the number of the surrounding vehicles. is the starting position of the vector, is the x-axis coordinate of the starting point of the vector, is the y-axis coordinate of the starting point of the vector, is the vector end position, is the x-axis coordinate of the vector's end point, is the y-coordinate of the vector's end point, It's speed. It is the direction, yes acceleration, onehot(i) is the unique hot encoding for the vehicle, target vehicle, and other vehicles, pad veh is zero-filled; finally, N surrounding vehicles T h The information at each moment is represented as follows: In the formula, Represents the vector features of the surrounding vehicle i at time t; A3. For local high-precision maps, all lane centerlines are divided into L non-overlapping segments, each consisting of D equidistant points; for lane segment l, adjacent points are connected into vectors; therefore, D equidistant points are represented as a set of D-1 vectors m l , the vector dimension is set to V dim , which is expressed as follows: In the formula, The superscript d∈[1,D-1] is the number of the equidistant point, and the subscript l∈[1,L] is the lane segment number. is the starting position of the vector, is the x-axis coordinate of the starting point of the vector, is the y-axis coordinate of the starting point of the vector, is the vector end position, is the x-axis coordinate of the vector's end point, is the y-coordinate of the vector's end point, is the vector yaw angle, is the turning direction, the turning direction includes left turn, right turn and straight Indicates whether it is at an intersection. Indicates whether there is a back neighbor, Indicates whether there is a front neighbor, idx is the index of the lane segment, pad map is zero-filled; finally, L lane segments consisting of D equidistant points are represented as follows: In the formula, Represents the vector feature of the dth equidistant point in lane segment l; B. Filter out illegal lane segments based on traffic rules and domain knowledge B1. Determine the seed lane segment based on traffic rules Calculate the positions of equally spaced points for each lane segment The current position of the target vehicle The minimum distance Δd, subscript l is the number of the lane segment, superscript d is the number of the equidistant point, subscript tar represents the target vehicle, and superscript 0 represents the current moment; Then, calculate the yaw angle at equidistant points in each lane segment The current direction of the target vehicle The minimum difference Δy is expressed as follows: Where dist(·) is the function for calculating the Euclidean distance, and min(·) is the function for taking the minimum value; In the first stage of determining the seed lane segment, the condition Δd < τ will be satisfied d And Δy<τ y The lane segment of is marked as a candidate seed lane segment, where τ d is the set distance threshold, τ y is the set orientation difference threshold; in the second stage of determining the seed lane segment, according to the traffic rules and lane topology, the candidate seed lane segment closest to the current position of the target vehicle in each lane is regarded as the seed lane segment Seed lane , the seed lane segment comes from the current lane or the adjacent lane; B2. Searching for legal lane segments based on domain knowledge First, according to the current state of the target vehicle and kinematic knowledge, the forward search depth τ is determined succ and the backward search depth τ prec , which is expressed as follows: t prec =d Where, T f is the predicted duration, The current speed of the target vehicle, The current acceleration of the target vehicle, 0<δ<500 is a fixed constant; Then, the seed lane segment Seed determined in step B1 is used lane As the search starting point, obtain the legal lane segment based on the breadth-first search algorithm It is expressed as follows: In the formula, BFS(·) is the breadth-first search algorithm, Link succ is the forward adjacency relationship of the lane, Link prec is the lane rearward adjacency relationship, It is the forward legal lane segment, It is a rearward legal lane section. is all legal lane segments; finally, step A3 preprocessing is used C. Vehicle and map information encoding and social interaction capture C1. Based on the unified representation of vehicle history information and high-precision map information, three encoders with the same structure and different weights, which are composed of gated multilayer perceptrons (gMLP), are used to encode the vector sets represented by steps A2, A3 and B2 respectively. and Extract vehicle features f v 、All lane segment features f m and legal lane segment characteristics It is expressed as follows: Where VehicleNet(·) is the vehicle feature encoder, W v is the weight for extracting vehicle features, MapNet(·) is the map feature encoder, and W m is the weight for extracting features of all lane segments, is the weight for extracting the features of legal lane segments; C2, vehicle feature f extracted based on step C1 v 、All lane segment features f m and legal lane segment characteristics The vehicle-vehicle interactor composed of self-attention mechanism is used to capture the vehicle-vehicle interaction features f v2v , a vehicle-map interactor consisting of a cross-attention mechanism is used to capture the vehicle-map interaction features f v2m , a map-map interactor consisting of a cross-attention mechanism is used to capture map-map interaction features It is expressed as follows: f v2v =SA(f v ,f v ) f v2m =CA(f v2v ,f m ) Where, SA(·) is the self-attention mechanism module, CA(·) is the cross-attention mechanism module; D. Multi-stage target prediction and target-guided decoding D1, based on the fact that the vehicle plans multiple intermediate goals in advance while moving to ensure safe arrival at the destination, the legal lane segment features are used Vehicle-map interaction features v2m and map-map interaction features Estimate the target lane segment of group K in stage J from the legal lane segment It is expressed as follows: In the formula, φ cat (·) is the feature concatenation operation, is a multi-layer perceptron, or MLP, softmax(·) is the softmax function, and score j is the probability score table of the jth stage, argtop(·,k) is the index function for taking the top k values, and index(·) is the function for taking values ​​by index; D2, following the paradigm of the target guidance method, using the multi-stage target lane segment guidance and constrained trajectory prediction of step D1; First, to promote the consistency and stability of the prediction scene, the multi-stage target lane segments are integrated through the cross-attention mechanism. and vehicle-map interaction v2m Fusion is performed to obtain fusion features It is expressed as follows: Then, in order to prevent the predicted trajectory and the target lane segment from deviating from each other over time, a dynamic adaptive adjuster is constructed based on the learnable query, which is expressed as follows: Where Φ is the learnable query, ⊙ is the element-wise dot product operation, and Θ is the output feature of the dynamic adaptive adjuster; Finally, to enhance the deep understanding of scene information, the map-map interaction feature As the initial hidden state of the gated recurrent unit, GRU, and as one of the conditions for probability prediction; multiple trajectories are obtained by GRU decoding, and the relevant probabilities are obtained by MLP decoding, which are expressed as follows: In the formula, Υ k is the predicted k-th trajectory, p k is the probability of the predicted kth track, and k is the number of the predicted track.

2. According to the multi-stage target-guided vehicle trajectory prediction method based on knowledge fusion according to claim 1, it is characterized by: The breadth-first search algorithm in step B2 is replaced by a depth-first search algorithm or a Monte Carlo tree search algorithm.

Images