Low-Bandwidth and Compute-Bound RGB-D Planar Semantic SLAM

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual simultaneous localization and mapping (SLAM) is crucial for intelligent mobile robots.
• Point-cloud map representations in RGB-D SLAM systems face limitations in onboard computation and communication bandwidth.

Purpose of the Study:

• To propose techniques for reducing computation and bandwidth load in RGB-D SLAM by mapping point clouds to parametric models.
• To integrate semantic information using a convolutional neural network (CNN) to simplify environmental geometric complexity.

Main Methods:

• Mapping point clouds to parametric models, specifically planar surfaces.
• Utilizing a CNN to extract semantic information for object modeling.
• Developing novel compression algorithms for depth data and a method for fitting planes to RGB-D data.
• Extending maps with semantic information predicted from sparse geometries.

Main Results:

• Demonstrated significant savings in computational and bandwidth resources compared to state-of-the-art SLAM systems.
• Enabled efficient knowledge sharing for multi-agent systems and human-robot cooperation.
• Facilitated real-time odometry estimation and mapping using plane data.

Conclusions:

• The proposed approach effectively reduces computational and bandwidth demands in RGB-D SLAM.
• Integrating semantic information with parametric models enhances the efficiency of world knowledge representation for robots.
• This work paves the way for more capable multi-agent robotic systems and human-robot collaboration.