CN-121980532-A - Future interaction modeling track prediction method for complex traffic scene

CN121980532ACN 121980532 ACN121980532 ACN 121980532ACN-121980532-A

Abstract

The invention relates to a future interaction modeling track prediction method for a complex traffic scene, which is used for coding the historical motion state of each traffic participant based on the historical track and a scene map of a target vehicle and surrounding participants thereof, carrying out future mode projection and time expansion on the historical features to generate potential future features, then constructing future interaction association characterization according to the association degree of the target vehicle and the surrounding participants in a future feature space, screening out a key participant set with potential future interaction relation with the target vehicle, carrying out interaction fusion on the future features of the key participant set and the future features of the target vehicle to obtain enhanced features with future interaction perception capability, and outputting a predicted track result of the target traffic participant in a future period based on the enhanced features, thereby improving the instantaneity, the prediction precision and the scene adaptation capability of track prediction under the complex traffic scene.

Inventors

LIU HONGFEI
LI JUNYANG
LI PENGLONG
XIANG JUNYI
YAO XINYU
YANG SHEN
YANG LIN

Assignees

吉林大学

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. A future interaction modeling track prediction method facing complex traffic scenes is characterized by comprising the following steps: step 1, acquiring the motion state of traffic participants, processing the position and the course angle of each traffic participant by taking the position and the course angle of the last frame of a target vehicle as references, backfilling the processed data to obtain the rotated historical state information; step 2, processing the data obtained in the step 1 to obtain the interaction characteristics of the historical track and the map information ; Step 3, based on the data of each traffic participant in the step 2, using an expert network to process to obtain expert characteristics, using a gate control network to distribute expert weights for the expert network in different future modes to obtain mode characteristics of the traffic participant in the future modes, then coding, and finishing all traffic participants and all future modes to obtain a potential future characteristic set; Step 4, projecting the potential future characteristics of each traffic participant to a reference coordinate system of the target vehicle, combining the potential future characteristics with the potential future characteristics of the target vehicle to construct a future interaction relation degree, organizing the potential future characteristics into a future interaction relation set according to a future mode after normalization, screening traffic participants with the top rank, acquiring the potential future characteristics related to the traffic participants, splicing the potential future characteristics with the potential future characteristics of the target vehicle, processing the potential future characteristics by a multi-head self-attention mechanism to acquire the future interaction characteristics in a certain future mode, and acquiring the future interaction characteristics after all data are spliced ; And 5, splicing the characteristics of the step 2 and the step 4, and processing the characteristics by a multi-head self-attention mechanism and a multi-layer perceptron to obtain the track of the target vehicle.
2. The future interaction modeling track prediction method for the complex traffic scene according to claim 1, wherein in step 1, firstly, the global position of the traffic participant at any historical moment is subjected to translational change, then the heading angle of the target vehicle at the last frame of the historical time step is used as the reference direction of the local coordinate system, the translated coordinates are subjected to rotational transformation to obtain the relative position of the traffic participant under the local coordinate system of the target vehicle, and then the rotational position information and the relative heading of the traffic participant relative to the target vehicle are backfilled to the historical state information to obtain the rotational historical state information.
3. The future interaction modeling track prediction method for complex traffic scenes according to claim 1, wherein step 2 uses a multi-layer perceptron to perform high-dimensional mapping on the rotated historical state information, and then obtains the historical information interaction characteristics after sequentially passing through a bidirectional LSTM encoder and a transducer encoder The high-precision map information is subjected to high-dimensional mapping by using a multi-layer perceptron, and then is encoded by using a polyline-based broken line encoder to obtain broken line map features Finally, the broken line map is characterized Features interacting with historical information Splicing, and capturing interaction between the target vehicle and the map by using a multi-head self-attention mechanism to obtain the interaction characteristics of the history track and the map information 。
4. The future interaction modeling trajectory prediction method for complex traffic scenarios of claim 1, wherein in step 3 the gating network consists of a linear mapping layer, relu activation functions and Softmax functions, the expert network consists of a multi-layer perceptron, two linear mapping layers, one Relu activation function, and the cyclic neural network is used to model characteristics of the ith traffic participant in the mth future mode Encoding to obtain potential future characteristics of the ith traffic participant in the mth future mode 。
5. The future interaction modeling trajectory prediction method for a complex traffic scene according to claim 1, wherein the step 4 specifically comprises the following steps: Step 4.1, projecting potential future characteristics of the traffic participant i into a reference coordinate system of the target vehicle by using MLP to obtain future interaction relation characteristics of the traffic participant i relative to the target vehicle; step 4.2. Based on potential future characteristics of the target vehicle in the mth future mode And future interaction relationship features of traffic participant i with respect to the target vehicle in the mth mode Constructing future interactive relation degree of association ; Step 4.3, normalizing the future interaction relation association degree of all traffic participants and the target vehicle in the scene to obtain the final future interaction relation representation ; Step 4.4. Organizing the set of future interaction associations of the target vehicle with all traffic participants in the mth mode , Representing the number of surrounding traffic participants; Step 4.5. Screening the traffic participants with the top 10 ranks to form a key set according to the future interaction association set obtained in the step 4.4 ; Step 4.6. According to the Key participant set Screening potential future features of the set of potential future features that are associated with the top 10 traffic participants; Step 4.7. After the potential future feature obtained in the step 4.6 is spliced with the potential future feature of the target vehicle, a multi-head self-attention mechanism is used for carrying out the future interaction between the target vehicle and the most relevant traffic participant, and the future interaction feature in the mth mode is obtained ; And 4.8. Splicing future interactions of the M modes to obtain future interaction characteristics.
6. The future interaction modeling trajectory prediction method for a complex traffic scene according to claim 1, wherein the future modes include a straight-through mode, a parking mode, a left-turn mode, a right-turn mode, a turning-around mode, and a lane-changing mode.
7. The future interaction modeling track prediction method for a complex traffic scene according to claim 3, wherein the multi-layer perceptron consists of two linear layers and one Relu activation function, the multi-head self-attention mechanism is 8 attention heads, and the number of attention layers is set to be 1.
8. The future interaction modeling trajectory prediction method for complex traffic scenarios of claim 4, wherein the pattern characteristics of the ith traffic participant in the mth future pattern in step 3 Wherein, the method comprises the steps of, Representing characteristics of historical track and map information interaction based on ith traffic participant The resulting kth expert feature is used to determine, Representing the gating weights assigned to the kth expert network by the ith traffic participant in the mth future mode, Representing the number of expert networks.
9. The method for predicting future interaction modeling trajectories for complex traffic scenarios of claim 5, wherein in step 4, the future interaction relationship characteristics of traffic participant i in the mth mode relative to the target vehicle The expression of (2) is: ; In the formula, Representing the relative position of traffic participant i in the target vehicle's local coordinate system, Representing the relative heading angle of the traffic participant i with respect to the target vehicle, the MLP is made up of two linear layers, one Relu activation function, Is a potential future feature of the ith traffic participant in the mth future mode.
10. The future interaction modeling track prediction method for complex traffic scenes according to claim 5, wherein the future interaction relation degree in step 4 is The expression is: ; In the formula, And Representing the query mapping matrix and the key mapping matrix respectively, As a dimension of the features, Representing transposed symbols.

Description

Future interaction modeling track prediction method for complex traffic scene Technical Field The invention belongs to the technical field of intelligent traffic and automatic driving, and particularly relates to a future interaction modeling track prediction method for a complex traffic scene. Background With the development of autopilot technology, motion prediction has become one of the key technologies in autopilot systems. The track prediction is mainly used for estimating the motion trend of the traffic participants in a future period according to the historical motion information, the surrounding scene information and the road topology constraint, so that prospective environment information support is provided for subsequent decision planning and risk avoidance. The existing track prediction method is used for modeling the interaction relation among traffic participants based on historical observation information, and generally adopts the processing modes of historical track coding, main body interaction modeling and future track decoding. Although the method can reflect the behavior association between the main bodies to a certain extent, the interaction modeling mainly depends on the observed historical information, and potential interactions possibly occurring at future moments are lack of effective characterization, so that the problems of track overlapping, inaccurate interaction judgment, unreasonable prediction results and the like are easy to occur in complex traffic scenes. To remedy the above-mentioned shortcomings, some methods further introduce a condition prediction mechanism, namely, assist the track prediction of the current subject by using the future states, future target points or future track estimation results of other traffic participants, for example, CN119068711a discloses a "merging area vehicle lane change track prediction system and method considering future space interaction", CN119068711a discloses a "vehicle track prediction method based on lane point future track offset auxiliary supervision". Although future interaction can be explicitly considered, the method generally depends on the prediction accuracy of the future state, is easy to generate error transfer problem, and has the defects of complex processing flow, limited real-time performance and the like. Therefore, it is needed to provide a multi-traffic-participant trajectory prediction method for complex traffic scenes, so as to realize potential future interaction relation modeling under the condition of not relying on explicit future trajectory priors, thereby improving accuracy, rationality and instantaneity of trajectory prediction. Disclosure of Invention In view of the problems that the existing track prediction method mainly relies on historical observation information to carry out interactive modeling, potential interactive relations at future moments are difficult to effectively characterize, and the condition prediction method based on the explicit future state has the problems of large error transfer, complex processing flow, insufficient instantaneity and the like, the invention provides the future interactive modeling track prediction method oriented to complex traffic scenes, the method oriented to the complex traffic scenes, an integrated processing framework of historical motion coding, potential future feature generation, future interaction object identification and track prediction output is constructed, the cooperative prediction of the future motion trend of multiple traffic participants is realized by modeling the future potential interaction relationship at the intermediate feature level, the method can directly mine potential future interaction relations from the historical observation information under the condition of not depending on the prior explicit future track, and improves the accuracy and rationality of track prediction of multiple traffic participants in a complex traffic scene. Compared with the existing processing method based on historical interaction or condition prediction, the method can effectively reduce dependence on future state prediction results, reduce error accumulation and propagation problems, reduce invalid interaction calculation through screening key interaction objects, and improve instantaneity, prediction precision and scene adaptation capability of track prediction in complex traffic scenes. In order to achieve the above purpose, the invention adopts the following technical scheme: A future interaction modeling track prediction method facing complex traffic scenes comprises the following steps: step 1, acquiring the motion state of traffic participants, processing the position and the course angle of each traffic participant by taking the position and the course angle of the last frame of a target vehicle as references, backfilling the processed data to obtain the rotated historical state information; step 2, processing the data obtained in the step 1 to obt