Lightweight 2.5D SLAM with Dynamic Map Refinement and Height-Aware Encoding for Resource-Constrained Indoor Robots

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Indoor mobile robots with sparse sensors face challenges in vertical perception and map artifacts.
• Existing SLAM systems struggle with dynamic environments and limited sensor data.

Purpose of the Study:

• To develop a lightweight 2.5D SLAM framework for low-cost indoor robots.
• To address limitations in vertical perception and dynamic artifacts in robot mapping.
• To enable compact, height-aware map representation for enhanced situational awareness.

Main Methods:

• Integrated a single-line laser distance sensor (LDS), Time-of-Flight (ToF), wheel odometry, and IMU for multi-sensor fusion.
• Employed an error-state Kalman filter (ESKF) for state estimation and NDT registration for map alignment.
• Implemented an offline dynamic refinement module with geometric clustering and filtering to remove transient artifacts.
• Introduced a 24-bit RGB occupancy encoding for compact, height-aware map storage.

Main Results:

• Achieved pose-consistent global map reconstruction with suppressed dynamic residual artifacts.
• Demonstrated effective suppression of transient artifacts while preserving stable structures in the map.
• Validated the framework's performance on public datasets and embedded hardware, showing efficient resource usage.
• Successfully implemented compact height-aware map encoding for practical deployment.

Conclusions:

• The proposed lightweight 2.5D SLAM framework enhances indoor robot perception and mapping capabilities.
• The system offers a practical solution for low-cost robots by combining sparse sensing with effective refinement.
• The height-aware map encoding facilitates improved navigation and scene understanding in complex indoor environments.