TY  - JOUR
AU  - Yu, Lian
AU  - Li, Dongsheng
AU  - Wang, Guoyan
AU  - Zhao, Fei
PY  - 2026
DA  - 2026/09/28
TI  - Design and Verification of a Centralized Multi-Agent Visual SLAM System Based on Cloud Native Architecture
JO  - Chinese Journal of Information Fusion
T2  - Chinese Journal of Information Fusion
JF  - Chinese Journal of Information Fusion
VL  - 3
IS  - 3
SP  - 209
EP  - 225
DO  - 10.62762/CJIF.2026.632794
UR  - https://www.icck.org/article/abs/CJIF.2026.632794
KW  - cloud-native architecture
KW  - multi-agent SLAM
KW  - map fusion
KW  - edge computing
AB  - To alleviate the computational and memory pressure faced by resource-constrained edge platforms in large-scale homogeneous multi-agent visual SLAM tasks, this paper designs and validates a centralized homogeneous multi-agent visual SLAM system based on a cloud-native architecture. The system adopts a cloud-edge functional decoupling strategy: real-time perception modules, such as visual odometry, keyframe extraction, and local tracking, are retained on the edge side, while computationally intensive tasks, including loop-closure detection, pose graph optimization, global map management, and multi-source map fusion, are centrally executed in the cloud. To accommodate bandwidth fluctuations, transmission latency, and short-term packet loss, the system establishes a ROS-based bidirectional asynchronous communication mechanism and combines keyframe-level compression, serialization, and ACK confirmation feedback to support reliable keyframe uploading. Meanwhile, an edge-side resource management workflow based on ``upload--acknowledge--release'' is designed to limit the growth of unacknowledged keyframe caches and improve long-term operational stability. For map fusion, the system adopts probabilistic occupancy-grid updating based on log-odds representation and, under homogeneous sensor conditions and constraints from globally optimized poses, projects local maps from multiple agents into a shared coordinate system for normalized equal-weight fusion. Experiments on public datasets, Gazebo collaborative simulations, baseline comparisons, communication-constrained tests, and long-horizon stress tests show that, under controlled experimental conditions, the proposed system reduces the edge-side load while maintaining localization accuracy and improves the stability of cloud-side map updating and long-term operation. Communication-disturbance experiments further analyze the influence of bandwidth limitation, transmission latency, random jitter, packet loss, and short-term outage on system performance, confirming the operational adaptability of the system under weak-network conditions.
SN  - 2998-3371
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Yu2026Design,
  author = {Lian Yu and Dongsheng Li and Guoyan Wang and Fei Zhao},
  title = {Design and Verification of a Centralized Multi-Agent Visual SLAM System Based on Cloud Native Architecture},
  journal = {Chinese Journal of Information Fusion},
  year = {2026},
  volume = {3},
  number = {3},
  pages = {209-225},
  doi = {10.62762/CJIF.2026.632794},
  url = {https://www.icck.org/article/abs/CJIF.2026.632794},
  abstract = {To alleviate the computational and memory pressure faced by resource-constrained edge platforms in large-scale homogeneous multi-agent visual SLAM tasks, this paper designs and validates a centralized homogeneous multi-agent visual SLAM system based on a cloud-native architecture. The system adopts a cloud-edge functional decoupling strategy: real-time perception modules, such as visual odometry, keyframe extraction, and local tracking, are retained on the edge side, while computationally intensive tasks, including loop-closure detection, pose graph optimization, global map management, and multi-source map fusion, are centrally executed in the cloud. To accommodate bandwidth fluctuations, transmission latency, and short-term packet loss, the system establishes a ROS-based bidirectional asynchronous communication mechanism and combines keyframe-level compression, serialization, and ACK confirmation feedback to support reliable keyframe uploading. Meanwhile, an edge-side resource management workflow based on ``upload--acknowledge--release'' is designed to limit the growth of unacknowledged keyframe caches and improve long-term operational stability. For map fusion, the system adopts probabilistic occupancy-grid updating based on log-odds representation and, under homogeneous sensor conditions and constraints from globally optimized poses, projects local maps from multiple agents into a shared coordinate system for normalized equal-weight fusion. Experiments on public datasets, Gazebo collaborative simulations, baseline comparisons, communication-constrained tests, and long-horizon stress tests show that, under controlled experimental conditions, the proposed system reduces the edge-side load while maintaining localization accuracy and improves the stability of cloud-side map updating and long-term operation. Communication-disturbance experiments further analyze the influence of bandwidth limitation, transmission latency, random jitter, packet loss, and short-term outage on system performance, confirming the operational adaptability of the system under weak-network conditions.},
  keywords = {cloud-native architecture, multi-agent SLAM, map fusion, edge computing},
  issn = {2998-3371},
  publisher = {Institute of Central Computation and Knowledge}
}