计算机科学
联想(心理学)
人工智能
数据关联
对象(语法)
关联规则学习
目标检测
模式识别(心理学)
计算机视觉
可视化
数据挖掘
雷达跟踪器
特征(语言学)
特征提取
数据建模
作者
Xuqi Yang,Tong Wang,Ning Zhang,Zhe Yu,He Chen
标识
DOI:10.1109/tgrs.2026.3724951
摘要
In multidrone cooperative perception scenarios, the core of multiview object association is to assign unique global identifiers (Global IDs) to the same physical target in different view images captured by K Drones. Existing multiview object association methods typically follow a bottom-up two-stage paradigm: first detecting objects in individual views, then associating detections across views. This paradigm has two drawbacks. First, it lacks global information from multiview context to supplement detection, making it difficult to handle situations such as occlusion. Second, the one-to-one correspondence in the association is enforced by rule-based constraints rather than being an inherent feature of the model, and such rule-based constraints cannot undergo backpropagation, which leads to cognitive conflict issues when it comes to multiview associations. Besides, the current two-view association datasets for drones covered up the importance of conflict-free association and global perception in the real-world multiview scenarios. To address these challenges, this article proposes a conflict-free cross-view object association framework for multidrone systems. Its core is a decoupled paradigm that maps single-instance features to multiview object queries, unifying multiview feature fusion (MVFF), global instance modeling, box prediction and conflict-free association within an end-to-end network. By modeling multiview features, the model naturally captures multiview contextual information and alleviates occlusion issues. Meanwhile, the uniqueness of instance features and the correspondence between bounding boxes and instances guarantee inherent one-to-one matching, which supports end-to-end backpropagation. To prove the effectiveness, four-view association datasets are constructed in the simulator, and two real-world sequences with three drone views are collected. Experimental results on the multidrone dataset demonstrate that the proposed framework achieves multidevice target association (MDA) score of 47.25%. Our method achieves favorable scalability in multiview scenarios. With the growth of view quantity, the unified cognition rate (UCR) undergoes a slight reduction caused by enhanced global consistency constraints, yet it consistently surpasses baseline approaches and adapts well to cloud platforms for large-scale multidrone systems. The code is available athttps://github.com/yyyyyxq/Instance-based-Association
科研通智能强力驱动
Strongly Powered by AbleSci AI