CN-121982877-A - Parallel iterative joint prediction planning method considering bidirectional interaction characteristics

CN121982877ACN 121982877 ACN121982877 ACN 121982877ACN-121982877-A

Abstract

The application relates to the technical field of track prediction and planning, in particular to a parallel iterative joint prediction planning method considering bidirectional interaction characteristics, wherein the method comprises the steps of constructing a bidirectional interaction modeling mechanism considering future track interaction; and decoding the final planning code and the final prediction code to obtain a multi-mode prediction track and a planning track of a planning time domain, and obtaining the final prediction track and the planning track according to the confidence level. Therefore, the problems that the coupling relation between prediction and planning is difficult to fully embody, the vehicle track cannot be timely adjusted according to traffic environment change, the planning conservation is easy to occur, the road traffic efficiency is influenced, and even traffic accidents are caused under complex dynamic scenes are solved because the related technology only establishes the unidirectional influence of prediction on the planning.

Inventors

LUO YUGONG
ZHAO XIANG
WANG JIANQIANG
LI KEQIANG
Ren hanxiao
HU YUNHAO
LI PENGFEI
Guan Shurui
YANG HAOYU

Assignees

清华大学

Dates

Publication Date: 20260505
Application Date: 20251217

Claims (8)

1. The parallel iterative joint prediction planning method considering the bidirectional interaction characteristics is characterized by comprising the following steps of: Dividing a planning time domain into a plurality of phases; In the multiple stages, performing interactive track prediction based on a bidirectional interactive modeling mechanism and interactive track planning based on the bidirectional interactive modeling mechanism in parallel according to future track coding iteration of the previous stage until preset conditions are met, so as to obtain final planning codes and final prediction codes respectively; And decoding the final planning code and the final prediction code to obtain a predicted track, a planning distance and corresponding confidence coefficient of the planning time domain, and mapping the planning distance to a transverse track to obtain the planning track of the planning time domain according to the predicted track, the confidence coefficient and the transverse track.
2. The method of claim 1, further comprising, prior to deriving the future track encoding of the previous stage: performing the interactive track prediction and the interactive track planning in parallel at starting step sections of the planning time domain to obtain an initial planning code and an initial prediction code; And splicing the initial planning code and the initial predictive code to obtain the future track code of the starting step sections.
3. The method of claim 1, wherein the interactive trajectory prediction based on the bi-directional interaction modeling mechanism comprises the steps of: performing a first cross-attention calculation on a scene code of a predicted target with a predicted query vector based on the bi-directional interaction modeling mechanism to obtain a current environmental semantic condition of the predicted target; Performing a second cross-attention calculation on the future trajectory codes of the previous stage with a predictive query vector to obtain a future intent condition of the agent; And splicing and time pooling interception are carried out on the current environment semantic condition and the future intention condition so as to obtain the predictive coding of the current stage.
4. A method according to claim 3, wherein the formula for the first cross-attention calculation is: , Wherein, the Representing the updated predictive query vector with the context information fused, A cross-attention calculation is indicated and, Representing the vector of the predictive query, Representing the scene code; the formula of the second cross attention calculation is: , Wherein, the A predictive query vector representing the updated fusion agent future intent, Representing the future track code.
5. The method of claim 4, wherein the formula for stitching is: , Wherein, the Represents an intermediate variable which is referred to as, Representing a first linear layer; Representing a first stitching operation; the formula of the time pooling interception is as follows: , Wherein, the A predictive coding is represented and is used to indicate, Representing the maximum pooling operation and, Representing the planned time domain of the j-th phase, The representation is truncated in the time dimension.
6. The method according to claim 1, wherein the interactive trajectory planning based on the bi-directional interaction modeling mechanism comprises the steps of: Determining a transverse candidate track of a planning target according to a preset rule strategy, and performing geometric feature coding on the transverse candidate track to obtain a transverse query vector of the planning target; Splicing and linearly changing the transverse query vector and the longitudinal query vector of the planning target, and performing self-attention mechanism calculation to obtain an initial planning query vector of the planning target; And carrying out interactive track planning on the initial planning query vector according to the bidirectional interactive modeling mechanism so as to obtain the planning code of the current stage.
7. The method of claim 6, wherein the formula for the linear change is: , Wherein, the Represents an intermediate variable which is referred to as, Representing a second linear layer; A second splicing operation is indicated and is shown, A lateral query is represented and, Representing a longitudinal query; the formula of the self-attention mechanism calculation is as follows: , Wherein, the Representing the initial planning query vector, Representing the self-attention mechanism calculation.
8. The method of claim 1, wherein the formula for predicting the trajectory is: , Wherein, the Representing a decoder; representing the predicted trajectory; representing the final predictive coding; the formula of the planning distance is as follows: , Wherein, the Representing the planned distance; Representing the final programming code; The confidence coefficient formula is: , Wherein, the Representing the confidence level.

Description

Parallel iterative joint prediction planning method considering bidirectional interaction characteristics Technical Field The application relates to the technical field of track prediction and planning, in particular to a parallel iterative joint prediction planning method considering bidirectional interaction characteristics. Background In an actual dynamic traffic environment, particularly under the complex working conditions of cities, such as areas of crossroads, ramps and the like, the vehicle flows are dense, the participation main bodies are various, the running order is relatively disordered, and various dynamic interaction behaviors are mutually interweaved and overlapped, so that the difficulty of planning the track of the automatic driving vehicle is remarkably increased. In the related art, a sequential method of "prediction-planning" is generally adopted, and the rationality of the planning result is improved by predicting the future states of surrounding traffic participants and then generating the planning track of the vehicle based on the prediction result. However, the related art only establishes a unidirectional influence of prediction on planning, and is difficult to fully consider a bidirectional coupling relation between prediction and planning, so that the safety and the passing efficiency of the auxiliary driving vehicle are still limited in a complex scene, and the problems of insufficient response of the auxiliary driving vehicle to environmental change, insufficient track adjustment, and reduced safety and reliability of vehicle driving decision are caused. Disclosure of Invention The application provides a parallel iterative joint prediction planning method considering bidirectional interaction characteristics, which aims to solve the problems that in the related technology, only the unidirectional influence of prediction on planning is established, so that the coupling relation between prediction and planning is difficult to fully embody, the track of a vehicle cannot be timely adjusted according to the change of traffic environment, and the problems of conservation of planning, influence on road traffic efficiency and even traffic accidents are easy to occur in a complex dynamic scene. An embodiment of a first aspect of the present application provides a parallel iterative joint prediction planning method considering bidirectional interaction characteristics, comprising the steps of dividing a planning time domain into a plurality of phases, performing interactive track prediction based on a bidirectional interaction modeling mechanism and interactive track planning based on the bidirectional interaction modeling mechanism in parallel according to future track coding iterations of a previous phase in the plurality of phases until preset conditions are met to obtain a final planning code and a final prediction code, decoding the final planning code and the final prediction code to obtain a predicted track, a planning distance and a corresponding confidence coefficient of the planning time domain, and mapping the planning distance to a transverse track to obtain the planning track of the planning time domain according to the predicted track, the confidence coefficient and the transverse track. Through the technical means, the embodiment of the application can obtain the predicted track of surrounding traffic participants and the planned track of the vehicle by executing the interactive track prediction and the interactive track planning in parallel, so as to fully integrate the interactive information among intelligent bodies and the environmental constraint, realize the accurate modeling of the interactive relation under the complex dynamic traffic scene, and guide the auxiliary driving or automatic driving vehicle to generate a safe and smooth final driving track, and simultaneously consider the traffic efficiency and the traffic rule. Optionally, in one embodiment of the present application, before obtaining the future track code of the previous stage, the method further comprises performing the interactive track prediction and the interactive track planning in parallel at starting step sections of the planning time domain to obtain an initial planning code and an initial prediction code, and splicing the initial planning code and the initial prediction code to obtain the future track code of the starting step sections. By the technical means, the embodiment of the application obtains the future track code through splicing the planning code and the predictive code, and can fuse the future state information of surrounding traffic participants and the planning track information of the vehicle, thereby forming high-dimensional characteristic representation considering interaction relation and environmental constraint, and simultaneously being used for subsequent track generation, decision optimization and safety evaluation, so that the auxiliary driving or automatic driving vehic