LTC-Mapping, Enhancing Long-Term Consistency of Object-Oriented Semantic Maps in Robotics

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Mobile robot operation relies on accurate environmental mapping.
• Existing object-oriented mapping methods struggle with instance duplication and dynamic scenes.
• Maintaining long-term map consistency is crucial for robot autonomy.

Purpose of the Study:

• To propose LTC-Mapping, a novel method for building long-term consistent object-oriented semantic maps.
• To address challenges of instance duplication and dynamic scenes in robot mapping.
• To enhance the reliability and accuracy of semantic maps for mobile robots.

Main Methods:

• Modeling detected objects using 3D bounding boxes and analyzing vertex visibility for occlusion detection.
• Augmenting geometric models with semantic information for object categorization.
• Employing data association and fusion techniques for temporal propagation of geometric and semantic data.
• Implementing a mechanism to remove objects from the map based on non-detection evidence to handle scene dynamics.

Main Results:

• LTC-Mapping demonstrates superior performance in modeling both geometric and semantic information of objects compared to state-of-the-art alternatives.
• The method effectively prevents instance duplication by analyzing object visibility and occlusions.
• LTC-Mapping successfully handles dynamic scenes, maintaining map accuracy over time.
• Experimental validation using the Robot@VirtualHome ecosystem confirms the method's effectiveness and suitability for online execution.

Conclusions:

• LTC-Mapping provides a robust solution for long-term consistent object-oriented semantic mapping in mobile robotics.
• The proposed approach enhances robot perception in complex and dynamic environments.
• LTC-Mapping offers a significant advancement for autonomous systems requiring accurate and persistent environmental understanding.