Conditional reconfiguration and time-of-day segment adaptive supervision trajectory generation method

Abstract

The present application belongs to the technical field of trajectory generation, and discloses a trajectory generation method with conditional recombination and time domain segmentation adaptive supervision. A representation recombination module is introduced after the conditional interaction module to adaptively recombine the trajectory representation fused with the scene condition, so that the trajectory representation can adapt to the importance difference of heterogeneous scene conditions in different traffic scenes, enhance the representation ability of the trajectory generation model for complex traffic scenes, and improve the scene consistency, interaction rationality and trajectory generation quality of the generated trajectory. Meanwhile, a time domain segmentation adaptive supervision mechanism is introduced in the training stage to divide the future time domain into multiple continuous time domain stages, and the trajectory errors at different future time steps are adaptively weighted and supervised to enhance the adaptive learning ability of the trajectory generation model for different stages of the future time domain.

Description

Technical Field

[0001] This invention relates to the field of trajectory generation technology, and in particular to a trajectory generation method with conditional recombination and temporal segmentation adaptive supervision. Background Technology

[0002] Trajectory generation is a crucial technology in intelligent systems such as autonomous driving and robotics. Its main task is to generate a motion trajectory for a future period based on the target object's state, historical trajectories, and scene conditions. In autonomous driving applications, scene conditions include surrounding traffic participants, road topology information, static object information, and navigation information. With the development of generative models, diffusion-based trajectory generation methods, due to their strong multimodal distribution modeling capabilities, are increasingly being applied to trajectory generation in complex dynamic scenarios.

[0003] Existing generative trajectory generation methods typically consist of two main parts: a scene encoder, which encodes traffic scene conditions such as the target object's state and historical trajectory, surrounding traffic participants, road topology information, static object information, and navigation information; and a trajectory generation decoder, which generates future trajectories under scene constraints. Taking the diffusion-based method "Yinan Zheng, Ruiming Liang, Kexin Zheng, Jinliang Zheng, Liyuan Mao, Jianxiong Li, Weihao Gu, Rui Ai, Shengbo Eben Li, Xianyuan Zhan, Jingjing Liu. Diffusion-Based Planning for Autonomous Driving with Flexible Guidance. ICLR 2025." as an example, it gradually restores the noisy trajectory to the target future trajectory during diffusion denoising, and can jointly model the autonomous vehicle's planned trajectory and the predicted trajectories of surrounding vehicles within the same generative framework.

[0004] First, while existing generative trajectory generation decoders can achieve conditional information fusion through the conditional interaction between scene conditions and trajectory representation, the relative importance of multi-source heterogeneous scene conditions such as road topology, surrounding traffic participants, static objects, and navigation intentions varies in trajectory generation. Furthermore, after conditional information fusion, trajectory representation is typically processed using a fixed static feature transformation method, making it difficult to adaptively reassemble the fused trajectory representation based on the interaction of different scene conditions within the current traffic scenario. Therefore, adaptive reassembly of the fused trajectory representation is necessary.

[0005] Second, different time stages of the future trajectory play different roles in the trajectory generation process. Existing training objectives usually aggregate the trajectory error over the entire future time domain in a unified or preset manner, which makes it difficult to reflect the differences in importance of different future time stages in the trajectory generation task. It is also difficult to adapt to the fact that near-term future trajectories directly affect safety, controllability, and interactive response, while far-term future trajectories have stronger uncertain time domain characteristics.

[0006] Therefore, a trajectory generation method with conditional recombination and temporal segmentation adaptive supervision needs to be proposed, which improves two aspects: trajectory representation processing after conditional fusion and future temporal supervision. The adaptive recombination capability of the trajectory representation after conditional fusion is enhanced at the trajectory denoising decoder level, and corresponding supervision weights are set for different continuous temporal stages at the supervision target level, thereby improving the scene consistency, interaction rationality and trajectory generation quality of the generated trajectory. Summary of the Invention

[0007] The purpose of this invention is to propose a trajectory generation method based on conditional recombination and temporal segmentation adaptive supervision. This method addresses the problems in existing generative trajectory generation methods, such as the varying relative importance of multi-source heterogeneous scene conditions, the use of fixed static transformations in the fused trajectory representation, the difficulty in adaptively recombinating the relationships between different scene conditions in the current traffic scenario, and the difficulty in reflecting the different trajectory generation characteristics at different future time stages due to the limitations of existing future temporal supervision. To address these issues, this invention introduces a representation recombination module after the conditional interaction module into the trajectory generation model to adaptively recombine the trajectory representation with fused scene conditions. Simultaneously, a temporal segmentation adaptive supervision mechanism is introduced into the training objective. By dividing the future time domain into multiple continuous time stages and assigning corresponding temporal segmentation adaptive supervision weights to different continuous time stages, the trajectory errors at different future time steps are weighted, enhancing the trajectory generation model's adaptive learning ability for different future time stages. This improves the scene consistency, interaction rationality, and trajectory generation quality of the generated trajectory.

[0008] The technical solution of the present invention is as follows: A trajectory generation method with conditional recombination and temporal segmentation adaptive supervision, comprising the following steps: Step 1: Obtain the current state of the target object, the historical trajectory of the target object, the information of surrounding traffic participants, the road topology information, the static object information, and the navigation information in the current target traffic scene as input to the target traffic scene, and obtain the noise information and diffusion time step information required for the diffusion generation process; Step 2: Transmit the target traffic scene input to the trajectory generation model; the trajectory generation model includes a scene encoder and a trajectory denoising decoder; the scene encoder encodes the target traffic scene input to obtain scene condition features; Construct the denoising trajectory based on the noise information; The trajectory to be denoised, the diffusion time step, and the scene condition features are input into the trajectory denoising decoder. After conditional interaction and representation reconstruction processing, a reconstructed trajectory representation is obtained, and a predicted clean future trajectory is output based on the reconstructed trajectory representation. Step 3: Construct temporal segmented adaptive supervision; During the training phase, the future time domain is divided into multiple continuous time domain stages, and corresponding temporal segmented adaptive supervision weights are set for different continuous time domain stages; The trajectory errors at different future time steps are weighted using the temporal segmented adaptive supervision weights to obtain the temporal segmented adaptive supervision loss, and the trajectory generation model is optimized based on the temporal segmented adaptive supervision loss.

[0009] The scene condition features are generated based on at least one of the following: the current state of the target object, the historical trajectory of the target object, information of surrounding traffic participants, road topology information, static object information, and navigation information.

[0010] The process of generating the trajectory to be denoised is as follows: During the training phase, the real future trajectory corresponding to the target traffic scene input is obtained, and the real future trajectory of the target object is compared with... The real future trajectories of the surrounding traffic participants are combined to form the real future joint trajectory: ; in, Indicates the number of future time steps. Indicates the number of traffic participants in the surrounding area. Represents the true future trajectory of the target object. Indicates the true future trajectory of surrounding traffic participants; Indicates the first M The true future trajectory of surrounding traffic participants; in the diffusion time step Based on the diffusion noise scheduling and noise information, the true future joint trajectory is determined. Perturbation is applied to obtain the noisy trajectory. , as the trajectory to be denoised; During the inference phase, an initial noise trajectory is generated based on the noise information, and the initial noise trajectory is used as the trajectory to be denoised.

[0011] The trajectory denoising decoder includes at least one decoding block, each decoding block including a trajectory internal dependency modeling module, a conditional interaction module, and a representation reconstructing module disposed after the conditional interaction module; feature mapping is performed on the trajectory to be denoised to obtain a trajectory representation to be denoised; trajectory internal dependency modeling is performed on the trajectory representation to be denoised through the trajectory internal dependency modeling module to obtain an internal dependency trajectory representation; conditional interaction is performed between the internal dependency trajectory representation and the scene conditional features through the conditional interaction module to obtain a conditional fused trajectory representation; the conditional fused trajectory representation is input into the representation reconstructing module to obtain a reconstructed trajectory representation.

[0012] The representation reconstruction module expands the conditionally fused trajectory representation into multiple residual representation streams, calculates input-related connection parameters based on the multiple residual representation streams, constructs branch inputs according to the input-related connection parameters, performs constrained inter-stream mixing on the multiple residual representation streams, performs branch transformation on the branch inputs to obtain branch outputs, writes the branch outputs back to the inter-stream mixed representation streams in a bounded manner, and aggregates the written-back representation streams to obtain the reconstructed trajectory representation.

[0013] The reorganization module includes a multi-stream expansion module, a connection parameter generation module, an input selection module, a constrained inter-stream mixing module, a branch transformation module, a bounded write-back module, and an aggregation module. set up , indicating the first The input in each decoding block represents the conditional fusion trajectory of the reassembly module. m Indicates the sequence number of the decoded block. The number of trajectory tokens is represented by D, and the feature dimension is represented by D; the conditional fusion trajectory representation Normalization is performed to obtain the normalized trajectory representation. ; in, This indicates a normalization operation. This represents the normalized trajectory representation; The normalized trajectory is represented as Expanded to multi-stream extension module Each residual represents the flow. , The residual represents the number of flows, and the first... Each residual is represented as a flow record. And each residual represents a flow With the normalized trajectory representation They have the same shape; Based on the multiple residual representations, the input relevant connection parameters are calculated by the connection parameter generation module; the input relevant connection parameters include a first connection parameter. Second connection parameters ; Will Divided into input selection parameters Inter-flow mixing parameters : Among them, input selection parameters Used to generate input selection weights corresponding to each residual representation stream, and to construct branch inputs based on the input selection weights; inter-stream mixing parameters This is used to generate constrained inter-flow mixing weights, and to redistribute information across multiple residual representation streams based on these constrained inter-flow mixing weights, resulting in multiple mixed representation streams. The input selection module is based on the input selection parameters. Generate input selection weights corresponding to each residual representation stream, and construct branch inputs by weighting and selecting information from multiple residual representation streams based on these input selection weights: in, This represents the activation function. Indicates the input selection weight. Indicates the first Each residual represents the input selection weight corresponding to the stream. Indicates branch input; The constrained inter-flow mixing module generates constrained inter-flow mixing weights based on the inter-flow mixing parameters, and redistributes information across multiple residual representation streams based on these weights to obtain multiple mixed representation streams: ; in, Represents the mixing weights between constrained flows. Indicates the first The residual indicates the flow to the first... A mixed weight representing the stream allocation information, Indicates the first A mixed representation stream, ; Perform a lightweight branch transformation on the branch input to obtain the branch output: in, This indicates the branch transformation operation corresponding to the branch transformation module. This represents the branch output obtained from the branch input through the branch transformation module; Based on the second connection parameters Generate bounded write-back weights and output the branch. Write back to each hybrid representation stream via the bounded write-back module Get the updated representation stream: ; in, Indicates the first The bounded write-back weights corresponding to the hybrid representation streams Indicates the first The updated representation stream is obtained by writing back the output of the hybrid representation stream through the branch; The aggregation module aggregates all updated representation streams to obtain the reconstructed trajectory representation: in, This represents the reorganization trajectory.

[0014] The predicted clean future trajectory is represented as follows: in, This represents the denoising mapping in the trajectory generation model. Indicates diffusion time step The corresponding denoised trajectory, Indicates the diffusion time step. Indicates scene condition features. This indicates a predicted clean future trajectory.

[0015] Step 3 involves constructing a temporal segmented adaptive supervised loss and optimizing the trajectory generation model, including: set up Indicates the first The training samples will be in the future time step The trajectory components of the target object that predicts the clean future trajectory. Let represent the true future trajectory state of the corresponding target object. Then, the single-step trajectory error is defined as: in, Indicates the batch sample index. Indicates the index of future time steps. Indicates the number of samples in the batch. Indicates the number of future time steps; Divide the future time domain into A continuous time-domain stage And set corresponding stage supervision weights for each continuous time-domain stage; for those that satisfy The next step in the future will be the first The stage supervision weights corresponding to each continuous time-domain stage are assigned as the temporal segmentation adaptive supervision weights for that future time step. ; Based on the time-domain segmented adaptive supervision weight We weight the single-step trajectory errors at each future time step to obtain the time-domain segmented adaptive supervision loss: ; in, This represents the time-domain piecewise adaptive supervised loss. Indicates future time steps The corresponding temporal segmented adaptive supervision weights, Indicates the first The training samples will be in the future time step The single-step trajectory error is calculated; the trajectory generation model is optimized using the temporal segmented adaptive supervised loss.

[0016] The beneficial effects of this invention are as follows: By deploying the representation reorganization module after the trajectory representation and scene conditions have interacted, this invention enables the trajectory representation after conditional fusion to undergo adaptive reorganization instead of being processed solely through fixed transformations, thereby enhancing the trajectory generation model's ability to represent complex traffic scenarios. This invention introduces a temporal segmentation adaptive supervision mechanism, dividing the future time domain into multiple continuous time domain stages and setting corresponding supervision weights for different continuous time domain stages. This allows the trajectory generation model to perform differentiated weighted supervision of trajectory errors at different future time steps during training, thereby enhancing its learning ability for different stages of the future time domain. This invention improves the trajectory generation model from two aspects: the generative trajectory decoding structure and the training supervision objective, thereby enhancing the scene consistency, interaction rationality, and trajectory generation quality of the generated trajectory. Attached Figure Description

[0017] Figure 1 A schematic diagram of the trajectory generation method with conditional recombination and temporal segmentation adaptive supervision; Figure 2 Comparison diagrams of trajectory generation results under different traffic scenarios: (a) Comparison algorithm under turning scenario; (b) Comparison algorithm under congested scenario; (c) Comparison algorithm under medium speed scenario; (d) This method under turning scenario; (e) This method under congested scenario; (f) This method under medium speed scenario. Figure 3 This is a schematic diagram of the reorganization module. Detailed Implementation

[0018] like Figure 1 As shown, a trajectory generation method with conditional recombination and temporal segmentation adaptive supervision is presented, and the specific steps are as follows.

[0019] Step 1: Obtain the target traffic scene input. Obtain the current state of the target object, the historical trajectory of the target object, the information of surrounding traffic participants, the road topology information, the static object information, and the navigation information in the current target traffic scene as the input of the target traffic scene. Obtain the noise information and diffusion time step information required for the diffusion generation process.

[0020] Step 2: Transmit the target traffic scene input to the trajectory generation model, which includes a scene encoder and a trajectory denoising decoder; use the scene encoder to encode the target traffic scene input to obtain scene condition features; The denoised trajectory is constructed based on the noise information; the denoised trajectory, diffusion time step and scene condition features are input to the trajectory denoising decoder, and the reconstructed trajectory representation is obtained through conditional interaction and representation reconstruction processing; and a predicted clean future trajectory is output based on the reconstructed trajectory representation.

[0021] Step 2.1: The scene condition features are generated based on at least one of the following: the current state of the target object, the historical trajectory of the target object, information of surrounding traffic participants, road topology information, static object information, and navigation information. The scene encoder is used to encode the target traffic scene input to obtain the scene condition features. , can be represented as: ; in, This represents the encoded scene condition features. This indicates the number of tokens representing scene condition features. Indicates the feature dimension.

[0022] Step 2.2: During the training phase, obtain the real future trajectory corresponding to the target traffic scene input, and compare the real future trajectory of the target object with... The combined real future trajectories of the surrounding traffic participants constitute a real future joint trajectory, which is denoted as: in, Indicates the number of future time steps. Indicates the number of traffic participants in the surrounding area. Represents the true future trajectory of the target object. Indicates the true future trajectory of surrounding traffic participants; Indicates the first M The actual future trajectory of surrounding traffic participants. In the diffusion time step. Based on the diffusion noise scheduling and noise information, the true future joint trajectory is determined. Perturbation is applied to obtain a noisy trajectory. , as the trajectory to be denoised.

[0023] In one specific implementation, it can be represented as: in, and Indicates diffusion time step The corresponding noise scheduling parameters, ε Indicates noise.

[0024] Step 2.3: The trajectory denoising decoder includes at least one decoding block, each decoding block including at least a trajectory internal dependency modeling module, a conditional interaction module, and a representation reconstructing module disposed after the conditional interaction module; the trajectory denoising decoder performs feature mapping on the trajectory to be denoised to obtain a representation of the trajectory to be denoised; the trajectory internal dependency modeling module performs trajectory internal dependency modeling on the representation of the trajectory to be denoised to obtain an internal dependency trajectory representation; the conditional interaction module performs conditional interaction between the internal dependency trajectory representation and the scene conditional features to obtain a conditional fusion trajectory representation; the conditional fusion trajectory representation is input into the representation reconstructing module to obtain a reconstructed trajectory representation. In one specific embodiment, the trajectory internal dependency modeling module adopts a self-attention module, the conditional interaction module adopts a cross-attention module; the representation reconstructing module is disposed after the cross-attention module.

[0025] Step 2.4: , indicating the first The input in each decoding block represents the conditional fusion trajectory of the reassembly module. Indicates the sequence number of the decoded block. Indicates the number of trajectory tokens. Representing the feature dimension; representing the conditional fusion trajectory. Normalization is performed to obtain the normalized trajectory representation. ; in, This indicates a normalization operation. This represents the normalized trajectory representation; Step 2.5: Represent the normalized trajectory Expand to Each residual represents the flow. , The residual represents the number of flows, and the first... Each residual is represented as a flow record. ， And each residual represents a flow. With the normalized trajectory representation They have the same shape; Step 2.6: Calculate the input correlation connection parameters based on the multiple residual representation streams; the input correlation connection parameters include the first connection parameter. Second connection parameters ; Will Divided into input selection parameters Inter-flow mixing parameters : Among them, input selection parameters Used to generate input selection weights corresponding to each residual representation stream, and to construct branch inputs based on the input selection weights; inter-stream mixing parameters This is used to generate constrained inter-flow mixing weights, and to redistribute information across multiple residual representation streams based on these constrained inter-flow mixing weights, resulting in multiple mixed representation streams. Step 2.7: The second connection parameters Used to generate bounded write-back weights; based on the input selection parameters Generate input selection weights corresponding to each residual representation stream, and construct branch inputs by weighting and selecting information from multiple residual representation streams based on these input selection weights: ; in, This represents the activation function. Indicates the input selection weight. Indicates the first Each residual represents the input selection weight corresponding to the stream. Indicates branch input; Step 2.8: Generate constrained inter-flow mixing weights based on the inter-flow mixing parameters, and redistribute information across multiple residual representation streams based on the constrained inter-flow mixing weights to obtain multiple mixed representation streams: ; in, Represents the mixing weights between constrained flows. Indicates the first The residual indicates the flow to the first... A mixed weight representing the stream allocation information, Indicates the first A mixed representation stream, ; Step 2.9: Perform a lightweight branch transformation on the branch input to obtain the branch output: in, This indicates the branch transformation operation corresponding to the branch transformation module. This represents the branch input constructed by the aforementioned input selection process. This represents the branch output obtained from the branch input through a branch transformation; Step 2.10: Based on the second connection parameters Generate bounded write-back weights and output the branch. Write back to each hybrid representation stream Get the updated representation stream: ; in, Indicates the first The bounded write-back weights corresponding to the hybrid representation streams Indicates the first The updated representation stream is obtained by writing back the output of the hybrid representation stream through the branch; Step 2.11: Aggregate all updated representation streams to obtain the reconstructed trajectory representation: ; in, This indicates the reorganization operation corresponding to the reorganization module. The reconstructed trajectory representation is characterized by maintaining the same number of trajectory tokens and feature dimensions in each updated representation stream. The aggregation module sums the corresponding positions of each updated representation stream, and the resulting reconstructed trajectory representation has the same number of trajectory tokens and feature dimensions as the normalized trajectory representation.

[0026] Step 2.12: Output a predicted clean future trajectory based on the reconstructed trajectory representation, wherein the predicted clean future trajectory is represented as: ; in, This represents the denoising mapping in the trajectory generation model. This represents the trajectory to be denoised corresponding to the diffusion time step. Indicates the diffusion time step. Indicates scene condition features. This indicates a predicted clean future trajectory.

[0027] Step 3: Construct temporal segmented adaptive supervision; During the training phase, the future time domain is divided into multiple continuous time domain stages, and corresponding temporal segmented adaptive supervision weights are set for different continuous time domain stages. The trajectory errors at different future time steps are weighted using the temporal segmented adaptive supervision weights to obtain the temporal segmented adaptive supervision loss, and the trajectory generation model is optimized based on the temporal segmented adaptive supervision loss. Step 3.1: Let Indicates the first The training samples will be in the future time step The trajectory components of the target object that predicts the clean future trajectory. Let represent the true future trajectory state of the corresponding target object. Then, the single-step trajectory error is defined as: in, Indicates the batch sample index. Indicates the index of future time steps. Indicates the number of samples in the batch. Indicates the number of future time steps; Step 3.2: Divide the future time domain into A continuous time-domain stage And set corresponding stage supervision weights for each continuous time-domain stage, for those that satisfy The next step in the future will be the first The stage supervision weights corresponding to each continuous time-domain stage are assigned as the temporal segmentation adaptive supervision weights for that future time step. ; Step 3.3: Adaptive Supervision Weights Based on the Time-Domain Segmentation We weight the single-step trajectory errors at each future time step to obtain the time-domain segmented adaptive supervision loss: ; in, This represents the time-domain piecewise adaptive supervised loss. Indicates future time steps The corresponding temporal segmented adaptive supervision weights, Indicates the first The training samples will be in the future time step The single-step trajectory error is addressed. The trajectory generation model is optimized using the temporal segmented adaptive supervised loss.

[0028] This invention does not merely apply generative models to trajectory generation, but improves upon them in two aspects: the trajectory generation structure and the training supervision objective. First, the invention places a representation reorganization module after the conditional interaction module. This module operates on the conditionally fused trajectory representation, which incorporates scene conditions such as road topology information, surrounding traffic participant information, static object information, and navigation information. This module achieves adaptive reorganization of the conditionally fused trajectory representation through operations such as multi-flow expansion, input selection, constrained flow mixing, branch transformation, bounded write-back, and aggregation, thereby enhancing the trajectory generation model's ability to represent different target traffic scene conditions. Second, this invention divides the future time domain into multiple continuous time domain stages and sets corresponding stage supervision weights for different continuous time domain stages. For different future time steps, the stage supervision weights corresponding to the continuous time domain stage to which it belongs are allocated as the temporal segmented adaptive supervision weights for that future time step, and these temporal segmented adaptive supervision weights are used to weight the trajectory errors at different future time steps. The aforementioned representation reorganization and temporal segmented adaptive supervision improve the trajectory generation model from two aspects: the trajectory generation structure and the training supervision objective.

Claims

1. A trajectory generation method with conditional recombination and temporal segmentation adaptive supervision, characterized in that, The steps include the following: Step 1: Obtain the current state of the target object, the historical trajectory of the target object, the information of surrounding traffic participants, the road topology information, the static object information, and the navigation information in the current target traffic scene as input to the target traffic scene, and obtain the noise information and diffusion time step information required for the diffusion generation process; Step 2: Transmit the target traffic scene input to the trajectory generation model; the trajectory generation model includes a scene encoder and a trajectory denoising decoder; the scene encoder encodes the target traffic scene input to obtain scene condition features; Construct the denoising trajectory based on the noise information; The trajectory to be denoised, the diffusion time step, and the scene condition features are input into the trajectory denoising decoder. After conditional interaction and representation reconstruction processing, a reconstructed trajectory representation is obtained, and a predicted clean future trajectory is output based on the reconstructed trajectory representation. Step 3: Construct temporal segmented adaptive supervision; During the training phase, the future time domain is divided into multiple continuous time domain stages, and corresponding temporal segmentation adaptive supervision weights are set for different continuous time domain stages. The trajectory error at different future time steps is weighted using the temporal segmented adaptive supervision weights to obtain the temporal segmented adaptive supervision loss, and the trajectory generation model is optimized based on the temporal segmented adaptive supervision loss.

2. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 1, characterized in that, The scene condition features are generated based on at least one of the following: the current state of the target object, the historical trajectory of the target object, information of surrounding traffic participants, road topology information, static object information, and navigation information.

3. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 1, characterized in that, The process of generating the trajectory to be denoised is as follows: During the training phase, the real future trajectory corresponding to the target traffic scene input is obtained, and the real future trajectory of the target object is compared with... The real future trajectories of the surrounding traffic participants are combined to form the real future joint trajectory: ； in, Indicates the number of future time steps. Indicates the number of traffic participants in the surrounding area. Represents the true future trajectory of the target object. Indicates the true future trajectory of surrounding traffic participants; Indicates the first M The true future trajectory of surrounding traffic participants; in the diffusion time step Based on the diffusion noise scheduling and noise information, the true future joint trajectory is determined. Perturbation is applied to obtain a noisy trajectory. , as the trajectory to be denoised; During the inference phase, an initial noise trajectory is generated based on the noise information, and the initial noise trajectory is used as the trajectory to be denoised.

4. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 1, characterized in that, The trajectory denoising decoder includes at least one decoding block, and each decoding block includes a trajectory internal dependency modeling module, a conditional interaction module, and a representation reconstructing module disposed after the conditional interaction module; The trajectory to be denoised is feature-mapped to obtain a representation of the trajectory to be denoised; the trajectory internal dependency modeling module is used to model the trajectory internal dependency of the trajectory to be denoised to obtain an internal dependency trajectory representation. The conditional interaction module performs conditional interaction between the internally dependent trajectory representation and the scene conditional features to obtain a conditional fusion trajectory representation; the conditional fusion trajectory representation is input into the representation recombination module to obtain a recombined trajectory representation.

5. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 4, characterized in that, The representation reconstruction module expands the conditionally fused trajectory representation into multiple residual representation streams, calculates input-related connection parameters based on the multiple residual representation streams, constructs branch inputs according to the input-related connection parameters, performs constrained inter-stream mixing on the multiple residual representation streams, performs branch transformation on the branch inputs to obtain branch outputs, writes the branch outputs back to the inter-stream mixed representation streams in a bounded manner, and aggregates the written-back representation streams to obtain the reconstructed trajectory representation.

6. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 5, characterized in that, The reorganization module includes a multi-stream expansion module, a connection parameter generation module, an input selection module, a constrained inter-stream mixing module, a branch transformation module, a bounded write-back module, and an aggregation module. set up , indicating the first The input in each decoding block represents the conditional fusion trajectory of the reassembly module. m Indicates the sequence number of the decoded block. D represents the number of trajectory tokens, and D represents the feature dimension. The conditional fusion trajectory representation Normalization is performed to obtain the normalized trajectory representation. ; in, This indicates a normalization operation. This represents the normalized trajectory representation; The normalized trajectory is represented as Expanded to multi-stream extension module Each residual represents the flow. , The residual represents the number of flows, and the first... Each residual is represented as a flow record. And each residual represents the flow With the normalized trajectory representation They have the same shape; Based on the multiple residual representations, the input relevant connection parameters are calculated by the connection parameter generation module; the input relevant connection parameters include a first connection parameter. Second connection parameters ; Will Divided into input selection parameters Inter-flow mixing parameters : Among them, input selection parameters Used to generate input selection weights corresponding to each residual representation stream, and to construct branch inputs based on the input selection weights; inter-stream mixing parameters Used to generate constrained inter-flow mixing weights, and to redistribute information across multiple residual representation streams based on the constrained inter-flow mixing weights, resulting in multiple mixed representation streams; The input selection module is based on the input selection parameters. Generate input selection weights corresponding to each residual representation stream, and construct branch inputs by weighting and selecting information from multiple residual representation streams based on these input selection weights: in, This represents the activation function. Indicates the input selection weight. Indicates the first Each residual represents the input selection weight corresponding to the stream. Indicates branch input; The constrained inter-flow mixing module generates constrained inter-flow mixing weights based on the inter-flow mixing parameters, and redistributes information across multiple residual representation streams based on these weights to obtain multiple mixed representation streams: ； in, Represents the mixing weights between constrained flows. Indicates the first The residual indicates the flow to the first... A mixed weight representing the stream allocation information, Indicates the first A mixed representation stream, ; Perform a lightweight branch transformation on the branch input to obtain the branch output: in, This indicates the branch transformation operation corresponding to the branch transformation module. This represents the branch output obtained from the branch input through the branch transformation module; Based on the second connection parameters Generate bounded write-back weights and output the branch. Write back to each hybrid representation stream via the bounded write-back module Get the updated representation stream: ； in, Indicates the first The bounded write-back weights corresponding to the hybrid representation streams Indicates the first The updated representation stream is obtained by writing back the output of the hybrid representation stream through the branch; The aggregation module aggregates all updated representation streams to obtain the reconstructed trajectory representation: in, This represents the reorganization trajectory.

7. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 1, characterized in that, The predicted clean future trajectory is represented as follows: in, This represents the denoising mapping in the trajectory generation model. Indicates diffusion time step The corresponding denoised trajectory, Indicates the diffusion time step. Indicates scene condition features. This indicates a predicted clean future trajectory.

8. The trajectory generation method with conditional recombination and temporal segmentation adaptive supervision according to claim 1, characterized in that, Step 3 involves constructing a temporal segmented adaptive supervised loss and optimizing the trajectory generation model, including: set up Indicates the first The training samples will be in the future time step The trajectory components of the target object that predicts the clean future trajectory. Let represent the true future trajectory state of the corresponding target object. Then, the single-step trajectory error is defined as: in, Indicates the batch sample index. Indicates the index of future time steps. Indicates the number of samples in the batch. Indicates the number of future time steps; Divide the future time domain into A continuous time-domain stage And set corresponding stage supervision weights for each continuous time-domain stage; for those that satisfy The next step in the future will be the first The stage supervision weights corresponding to each continuous time-domain stage are assigned as the temporal segmentation adaptive supervision weights for that future time step. ; Based on the time-domain segmented adaptive supervision weight We weight the single-step trajectory errors at each future time step to obtain the time-domain segmented adaptive supervision loss: ； in, This represents the time-domain piecewise adaptive supervised loss. Indicates future time steps The corresponding temporal segmented adaptive supervision weights, Indicates the first The training samples will be in the future time step The single-step trajectory error is calculated; the trajectory generation model is optimized using the temporal segmented adaptive supervised loss.

Images