CDJMP: an adaptive conditional diffusion model for multi-agent joint motion prediction in autonomous driving
推荐理由
存在代码/项目页线索，利于进入复现
核心判断
论文摘要（中文）
论文提出 CDJMP，用自适应条件扩散模型预测自动驾驶中的多车联合轨迹。两阶段流程先预测候选轨迹并自适应确定去噪步数，再用群体感知条件编码器建模车辆交互，生成连贯的多模态轨迹。
Inautonomousdrivingtrajectoryprediction,itisimportanttogeneratemulti-modaltrajectories.Asagenerative method,diffusionmodelhasbeenincreasinglyadoptedinthefieldoftrajectoryprediction.Inthispaper,we proposeCDJMP(ConditionalDiffusionmodel-basedJointMotionPrediction),anadaptiveconditionaldiffusion frameworkdesignedformulti-agenttrajectoryforecastinginautonomousdriving.However,whenconventional diffusionmodelsgeneratemulti-modaltrajectories,particularlyformulti-agentjointprediction,theyoftenfailto balancethediversityofgeneratedtrajectoriesandtheaccuracyofprediction.Atthesametime,diffusionmodels alsosufferfromtime-consuminginference.Toaddresstheseissues,CDJMPutilizesatwo-stageframeworkto generatediverseandhighlyaccuratetrajectories.Inthefirststage,aprobabilisticinitializerpredictsproposal trajectoriestogetherwithadaptivedenoisingsteps.Inthesecondstage,agroup-awareconditionalencoder capturesdynamicmulti-agentinteractionsandguidesthediffusionprocesstoproducecoherentmultimodal outcomes.ExperimentsontheINTERACTIONandArgoversedatasetsdemonstratethatCDJMPachieves state-of-the-artperformance,reducingminADEandminFDEbyupto9%and10%,respectively,onthe INTERACTIONdataset.AblationexperimentsdemonstratethatCDJMPcaneffectivelyreduceinferencetimewhile maintainingpredictionaccuracy.TheseresultshighlightthepotentialofCDJMPasanefficientandaccurate frameworkforreal-timemulti-agenttrajectorypredictioninautonomousdriving.
研究动机（中文总结）
多车联合预测需要同时保持轨迹多样性与精度，但常规扩散方法难以兼顾两者，而且迭代去噪带来较高推理耗时。
To overcome these limitations, we propose CDJMP, a Conditional Diffusion model based Joint Motion Prediction framework that integrates adaptive diffusion control and group-aware interaction modeling.
创新与贡献（中文总结）
把自适应扩散步数与群体交互建模结合，使扩散预测针对场景和交互关系调整，而不是固定步数地独立生成轨迹。
To overcome these limitations, we propose CDJMP, a Conditional Diffusion model based Joint Motion Prediction framework that integrates adaptive diffusion control and group-aware interaction modeling.
方法与证据
方法（中文总结）
第一阶段由概率初始化器生成候选轨迹并预测所需去噪步数；第二阶段的群体感知条件编码器提取动态多车交互信息，引导扩散过程生成联合多模态结果。
To overcome these limitations, we propose CDJMP, a Conditional Diffusion model based Joint Motion Prediction framework that integrates adaptive diffusion control and group-aware interaction modeling.
实验结果（中文总结）
在 INTERACTION 数据集上，摘要报告 minADE 和 minFDE 最多分别降低 9% 和 10%；在 INTERACTION 与 Argoverse 上优于所比较方法，并通过消融实验检验推理时间与精度的权衡。
Extensive experiments on the INTERACTION and Argoverse datasets demonstrate that our method achieves state-of-the-art performance on key metrics, including minADE, minFDE, and MR,significantlyoutperformingpriorapproachessuchas HyperMTP and LED, especiallyin highly interactive drivingscenarios.