Convolutional Neural Network-Based Robot Navigation Using Uncalibrated Spherical Images.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Traditional mobile robot navigation methods like SLAM and lane detection are computationally intensive and require extensive calibration.
• Existing approaches often struggle with high computational costs and demanding calibration processes.

Purpose of the Study:

• To propose a lightweight robot navigation framework using uncalibrated spherical images.
• To enhance computational efficiency and simplify orientation estimation and path prediction.

Main Methods:

• Decomposing the navigation problem into a series of classification tasks.
• Utilizing a spherical camera for 360° fisheye panorama capture to generate sufficient training samples.
• Implementing an end-to-end Convolutional Neural Network (CNN) trained on the Spherical-Navi dataset.

Main Results:

• The proposed CNN accurately predicts potential path directions from single, uncalibrated spherical images.
• The framework demonstrates superior performance compared to competing methods in realistic applications.
• Achieved high confidence levels in path direction predictions.

Conclusions:

• The developed lightweight framework offers an efficient solution for vision-based mobile robot navigation.
• The use of spherical images and a CNN-based classification approach effectively addresses limitations of previous methods.
• The proposed system shows significant promise for real-world robotic applications.