SLAMM: Visual monocular SLAM with continuous mapping using multiple maps.

Area of Science:

• Robotics
• Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Real-world robotic applications require robust localization and mapping systems.
• Existing SLAM systems often fail due to tracking disruptions from corrupted frames or sensor malfunctions.
• Maintaining continuous mapping and preserving information during such failures is a significant challenge.

Purpose of the Study:

• To present Simultaneous Localization and Multi-Mapping (SLAMM), a novel system designed for continuous mapping and information preservation.
• To develop a flexible SLAM system capable of handling tracking failures and operating with single or multiple robots.
• To improve the reliability and efficiency of SLAM in real-world scenarios.

Main Methods:

• SLAMM generates new maps upon tracking failure and merges them during loop closure events for single robots.
• For multi-robot scenarios, SLAMM merges maps without prior knowledge of relative poses, ensuring flexibility.
• The system operates in real-time at frame-rate speed, processing data efficiently.

Main Results:

• The proposed SLAMM approach demonstrated superior performance compared to state-of-the-art methods on KITTI and TUM RGB-D datasets.
• Mean tracking time was approximately 22 milliseconds, indicating real-time capability.
• Initialization was twice as fast as ORB-SLAM, and information preservation increased by up to 90%.

Conclusions:

• SLAMM provides a robust and flexible solution for continuous mapping and information preservation in the presence of tracking failures.
• The system's real-time performance, faster initialization, and enhanced data preservation make it suitable for demanding real-world applications.
• The open-source release of SLAMM and its framework facilitates community adoption and further development.