Multi-Agent Collaborative Navigation Based on Cross-View Matching

Collaboration between intelligent agents have gained significant research attention due to their practicality across various scenarios. This work addresses the problem of collaborative navigation for intelligent agents and proposes a distributed estimation framework centered on cross-view matching. To obtain cross-view observations, by integrating object detection algorithms with depth cameras, this work proposes a multi-view matching strategy that combines semantic and positional information. The proposed fusion framework employs a Error-State Kalman Filter (ESKF) to fuse ranging data, visual-inertial odometry (VIO) information, and cross-view observations. It further establishes cross-view geometric constraints by leveraging environmental features and targets that are visible to multiple agents. Compared to traditional integrated navigation methods, experimental results demonstrate that introducing co-visibility constraints effectively enhances pose estimation accuracy while ensuring both precision and robustness in GNSS-denied environments, and highlight the potential of the proposed approach as a lightweight and reliable navigation solution in GNSS-denied environments. This work provides technical support for vision-centric collaborative autonomous systems.