DeHCP:面向可扩展开放异构协同感知的解耦框架OA
DeHCP:A Decoupled Framework for Scalable Open Heterogeneous Collaborative Perception
现有开放式异构协同感知方法大多采用统一映射策略,将来自不同传感器和模型架构的特征投影至同一语义空间.然而,该过程不可避免地引起信息损失,并削弱各模态所具有的特有细节.针对这一问题,本文提出一种解耦式异构协同感知(Decoupled heterogeneous collaborative perception,DeHCP)框架,将中间特征空间划分为共享分支和特定分支:共享分支用于提取与模态无关的公共表征,特定分支则在可学习模态嵌入的引导下保留模态特有信息.进一步地,本文设计动态门控聚合机制,依据全局语义上下文和自车模态身份,自适应融合模态特定特征与公共表征.同时,提出一种结合反向对齐的解耦监督训练策略,使新加入的异构智能体仅需更新本地编码器和轻量化适配模块即可接入既有协同系统,无需对整个协同网络进行联合重训练.在协同感知基准数据集上的大量实验表明,DeHCP 在保持开放式异构协同感知良好可扩展性的同时,进一步提升了三维目标检测精度,为构建具备跨模态兼容能力的高可扩展自动驾驶协同系统提供了有效方案.代 码 已 开 源:https://github.com/jurui-cloud/DeHCP.
The existing open heterogeneous collaborative perception systems typically align diverse features to a fixed semantic space.This alignment process inevitably leads to information loss and suppresses modality-specific characteristics.To address this limitation,this paper introduces a novel decoupled heterogeneous collaborative perception(DeHCP)framework,which decouples the intermediate feature space into a shared branch and a specific branch.The shared branch extracts a modality-agnostic common representation,while the specific branch preserves modality-specific characteristics conditioned on learnable modality embeddings.A dynamic gate aggregation mechanism adaptively integrates these specific features with the common representation,based on the global semantic context and ego-agent identity.Furthermore,a decoupled supervision strategy with backward alignment enables the integration of unseen heterogeneous agents by updating only local encoders and lightweight adapters,thereby avoiding collective retraining.Extensive experiments on collaborative benchmarks demonstrate that DeHCP achieves improved three-dimensional object detection accuracy while maintaining the extensibility of open heterogeneous collaborative perception through backward alignment.The proposed approach supports the development of highly scalable autonomous driving systems.Code is available at https://github.com/jurui-cloud/DeHCP.
孙远;曹俱睿;刘宇颖;张栋;李佐勇;杜少毅
西安交通大学人工智能与机器人研究所人机混合增强智能全国重点实验室,西安 710049,中国西安交通大学人工智能与机器人研究所人机混合增强智能全国重点实验室,西安 710049,中国西安交通大学人工智能与机器人研究所人机混合增强智能全国重点实验室,西安 710049,中国西安交通大学人工智能与机器人研究所人机混合增强智能全国重点实验室,西安 710049,中国闽江大学计算机与大数据学院福建省信息处理与智能控制重点实验室,福州 350121,中国西安交通大学人工智能与机器人研究所人机混合增强智能全国重点实验室,西安 710049,中国
信息技术与安全科学
协同感知异构智能体传感器融合三维目标检测自动驾驶车辆
collaborative perceptionheterogeneous agentsensor fusion3D object detectionautonomous vehicles
《南京航空航天大学学报(英文版)》 2026 (4)
517-528,12
This work was supported by the Na-tional Key Research and Development Program of China(No.2025ZD0124204),the National Natural Science Foun-dation of China(No.U24A20252),the Fundamental and In-terdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China(No.JYB2025XDXM504),the Fu-zhou Artificial Intelligence"List-Bidding&Task-Commissioning"Project(No.2025-ZD-038),the Natural Sci-ence Basic Research Plan in Shaanxi Province of China(No.2026JC-YBQN-0779),and the Fundamental Research Fund for the Central Universities(No.xzy012026035).
评论