[Paper Review] A Collaborative Visual SLAM Framework for Service Robots

Mingjun Ouyang, Xuesong Shi+6 more|arXiv (Cornell University)|Feb 5, 2021

Robotics and Sensor-Based Localization24 references4 citations

TL;DR

This paper proposes a collaborative visual SLAM framework for service robots using an edge server to maintain a centralized map database, enabling real-time, low-overhead information sharing through efficient landmark retrieval. By allowing robots to query only view-predicted landmarks and perform lightweight updates, the system achieves scalable, accurate localization and map merging across multiple robots and cameras with minimal communication and computation costs.

ABSTRACT

We present a collaborative visual simultaneous localization and mapping (SLAM) framework for service robots. With an edge server maintaining a map database and performing global optimization, each robot can register to an existing map, update the map, or build new maps, all with a unified interface and low computation and memory cost. We design an elegant communication pipeline to enable real-time information sharing between robots. With a novel landmark organization and retrieval method on the server, each robot can acquire landmarks predicted to be in its view, to augment its local map. The framework is general enough to support both RGB-D and monocular cameras, as well as robots with multiple cameras, taking the rigid constraints between cameras into consideration. The proposed framework has been fully implemented and verified with public datasets and live experiments.

Motivation & Objective

• To enable multiple service robots to collaboratively build and reuse a shared, up-to-date map in dynamic environments.
• To reduce communication and computation overhead by transmitting only keyframe features and optimized landmarks instead of raw images.
• To support heterogeneous camera setups, including RGB-D, monocular, and multi-camera robots, with rigid inter-camera constraints.
• To design a scalable client-server architecture that maintains map consistency while preserving client autonomy during intermittent connectivity.
• To enable real-time landmark retrieval from the server based on predicted camera views, minimizing bandwidth usage.

Proposed method

• The system uses a client-server architecture where each robot maintains a lightweight local map for real-time tracking, while the edge server stores and optimizes a global map.
• Robots communicate only feature points and optimized landmarks, reducing bandwidth and server-side computation.
• A grid-based landmark indexing system enables fast, real-time retrieval of landmarks likely to appear in each robot’s current field of view.
• The server performs loop detection, map merging, and global pose graph optimization to improve accuracy and consistency.
• Multi-camera robots are modeled as multiple independent clients with inter-camera constraints enforced during global optimization.
• The framework is built on OpenVSLAM and supports both monocular and RGB-D cameras, with seamless integration of multi-camera data.

Experimental results

Research questions

• RQ1 How can multiple service robots efficiently share visual SLAM data with minimal communication overhead?
• RQ2 How can a centralized edge server maintain a consistent, globally optimized map while supporting real-time client updates?
• RQ3 What is an effective method to retrieve only relevant landmarks for a robot’s current view, minimizing bandwidth usage?
• RQ4 How can the system support heterogeneous camera types and multi-camera robots with rigid inter-camera constraints?
• RQ5 To what extent does collaborative map merging improve localization accuracy compared to single-robot SLAM?

Key findings

• The system achieved slightly better absolute trajectory error (ATE) than OpenVSLAM: 2.51m vs. 2.53m on the market1-1 sequence, and 3.34m vs. 3.70m on market1-2, indicating improved accuracy through collaborative loop closure.
• The average uplink communication cost was 0.5 MB/s per client, and downlink was 0.1 MB/s, demonstrating efficient data transfer with minimal bandwidth usage.
• Server-side landmark retrieval took only 1.7 ms on average, confirming the efficiency of the grid-based search method.
• Map merging and pose graph optimization took 0.60s and 1.12s per cycle, respectively, showing acceptable latency for real-time operation.
• The system maintained stable performance with up to eight clients on a quad-core server, with CPU usage between 100%-150%, indicating scalability.
• The final map from five sessions with a real multi-camera robot covered all explored areas and aligned well with the office floorplan, even after tracking failures and viewpoint changes.