Self-Supervised Sidewalk Perception Using Fast Video Semantic Segmentation for Robotic Wheelchairs in Smart Mobility.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Real-time semantic segmentation is vital for autonomous navigation in complex urban environments.
• Existing video segmentation methods require further adaptation for practical deployment.

Purpose of the Study:

• To develop a robust and efficient visual perception technique for robotic wheelchairs operating in urban sidewalk settings.
• To enable real-time segmentation with lightweight flow estimations and reliable feature extraction.

Main Methods:

• Utilized a novel approach based on recent video segmentation trends.
• Developed a visual perception technique tailored for urban sidewalk environments.
• Trained and validated the approach using a collection of synthetic scenes.

Main Results:

• Achieved improved prediction accuracy on the benchmark dataset.
• Demonstrated tolerable loss of speed with no additional overhead.
• The method proved effective for real-time segmentation in urban sidewalk scenarios.

Conclusions:

• The proposed technique offers a practical solution for real-time video semantic segmentation in challenging domains.
• It serves as a valuable reference for transferring and developing perception algorithms across diverse cross-domain applications.
• The method reduces downtime and enhances the feasibility of autonomous navigation systems.