高级检索

动态多视角协同感知综述

Survey of Dynamic Multi-View Collaborative Perception

  • 摘要: 动态多视角协同感知旨在解决多移动平台、多异构视角与持续演化场景下的全局感知问题,其核心挑战在于如何在相对位姿时变、目标可见性跳变以及协同关系不稳定的条件下,实现跨视角空间对齐、身份一致性推理与高层语义理解。本文以“几何动态—观测动态—协同动态”为统一分析主线,对动态多视角协同感知的代表性研究进行系统综述。针对几何动态,本文梳理动态位姿估计、互补视角连接、共享鸟瞰表示、动态神经辐射场与四维高斯泼溅等统一时空表示方法;针对观测动态,分析跨视角关联与跟踪、空地协同感知、车路协同感知以及第一人称—第三人称视角协同理解等研究方向;针对协同动态,讨论时间异步校准、拓扑时变系统与通信受限条件下的分布式协同机制。在此基础上,本文进一步总结能够反映几何动态、观测动态与协同动态特征的代表性公开数据集、核心评测指标与公开基准结果,分析不同技术路线的适用条件、性能优势与失效边界。最后,本文从统一时空表示与在线空间对齐、开放类别与跨域鲁棒关联、异步拓扑自适应协同机制,以及面向真实部署的系统级评测基准四个方面讨论未来发展方向。

     

    Abstract: Dynamic multi-view collaborative perception aims to address the problem of global perception in scenarios involving multiple mobile platforms, heterogeneous viewpoints, and continuously evolving environments. Its core challenge lies in achieving cross-view spatial alignment, identity-consistent reasoning, and high-level semantic understanding under conditions of time-varying relative poses, abrupt changes in target visibility, and unstable collaborative relationships. This paper takes “geometric dynamics, observation dynamics, and collaborative dynamics” as a unified analytical thread and provides a systematic survey of representative studies on dynamic multi-view collaborative perception. For geometric dynamics, this paper reviews dynamic pose estimation, complementary-view connection, shared bird’s-eye-view representation, and unified spatiotemporal representation methods such as dynamic neural radiance fields and four-dimensional Gaussian splatting. For observation dynamics, this paper analyzes research directions including cross-view association and tracking, air-ground collaborative perception, vehicle-infrastructure collaborative perception, and collaborative understanding between first-person and third-person views. For collaborative dynamics, this paper discusses temporal asynchronous calibration, topology-varying systems, and distributed collaborative mechanisms under communication constraints. On this basis, this paper further summarizes representative public datasets, core evaluation metrics, and public benchmark results that reflect the characteristics of geometric dynamics, observation dynamics, and collaborative dynamics, and analyzes the applicable conditions, performance advantages, and failure boundaries of different technical routes. Finally, this paper discusses future research directions from four aspects: unified spatiotemporal representation and online spatial alignment, open-category and cross-domain robust association, asynchronous topology-adaptive collaborative mechanisms, and system-level evaluation benchmarks for real-world deployment.

     

/

返回文章
返回