Robust and Fast Scene Recognition in Robotics Through the Automatic Identification of Meaningful Images

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Scene recognition is crucial in robotics but is often view-dependent and prone to bias.
• Existing methods struggle with generalization across different viewpoints.
• This research addresses the challenge of identifying optimal views for scene recognition.

Purpose of the Study:

• To develop a system that filters noisy images and retains informative views for scene recognition.
• To improve the reliability of scene identification in robotic applications.
• To validate a novel approach for view selection in scene recognition.

Main Methods:

• A heuristic metric based on Harris 3D key point detection in 3D meshes.
• A machine learning model combining a Minimum Spanning Tree and a Support Vector Machine (SVM).
• Extensive experiments using public and custom image databases.

Main Results:

• Identification of efficient visual descriptors for scene recognition.
• Analysis of the heuristic metric's alignment with human relevance criteria.
• Experimental validation of the proposed filtering system.

Conclusions:

• The developed system effectively filters irrelevant images, retaining optimal views for scene recognition.
• The heuristic metric shows promise in mimicking human judgment for view relevance.
• This approach enhances the robustness of scene recognition in robotics.