Effective Free-Driving Region Detection for Mobile Robots by Uncertainty Estimation Using RGB-D Data

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous mobile robots require accurate segmentation of drivable areas and obstacles for safe navigation.
• Current deep learning methods struggle with novel objects not present in training data, and require extensive manual labeling.
• This limitation hinders the performance and scalability of autonomous systems in complex, real-world environments.

Purpose of the Study:

• To propose a self-supervised learning method for segmenting drivable areas and road anomalies.
• To develop an efficient framework for automatic generation of segmentation labels, reducing reliance on manual annotation.
• To enable robust autonomous navigation by addressing challenges posed by unfamiliar objects and data scarcity.

Main Methods:

• Introduction of the Automatic Generating Segmentation Label (AGSL) framework.
• AGSL automatically generates segmentation labels by identifying image dissimilarities and localizing obstacles in disparity maps.
• Training a semantic segmentation network on RGB-D datasets using self-generated AGSL labels to create a pre-trained model.

Main Results:

• The AGSL framework demonstrated high performance in labeling evaluation.
• The pre-trained model showed promising results for real-time segmentation applications on mobile robots.
• The method effectively addresses the limitations of supervised learning in autonomous navigation.

Conclusions:

• The proposed self-supervised learning approach, utilizing the AGSL framework, significantly enhances the segmentation of drivable areas and obstacles.
• This method reduces the dependency on large, manually labeled datasets, making autonomous robot development more efficient.
• The findings pave the way for more reliable and adaptable autonomous navigation systems capable of handling diverse and unencountered scenarios.