Autonomous Thermal Vision Robotic System for Victims Recognition in Search and Rescue Missions.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Post-disaster environments (PDE) present significant challenges for victim identification due to low light and debris.
• Traditional methods like RGB imaging and trained dogs have limitations in these conditions.
• Advancements in neural networks and thermal imaging offer new possibilities for robotic search and rescue.

Purpose of the Study:

• To develop and validate a novel automatic method for victim identification in PDEs using thermal imaging and Convolutional Neural Networks (CNNs).
• To implement this method on a quadruped robotic system for data capture and processing.
• To compare the proposed method's efficiency against CNN-based RGB image processing.

Main Methods:

• A quadruped robot equipped with thermal and RGB cameras was used for data acquisition in simulated PDEs.
• Thermal images were processed using Convolutional Neural Networks (CNNs), including Fast R-CNN, SSD, and YOLO models.
• The Robot Operating System (ROS) was employed for robot control and data processing.

Main Results:

• The proposed thermal imaging and CNN-based method demonstrated high efficiency in victim identification within simulated PDEs.
• The system achieved an accuracy greater than 90% in detecting victims under various challenging conditions.
• The method proved more effective than conventional CNN-based RGB image processing in post-disaster scenarios.

Conclusions:

• The developed automatic method using thermal imaging and CNNs is highly effective for victim identification in post-disaster environments.
• The quadruped robotic system provides a viable platform for data capture and real-time analysis in search and rescue operations.
• This technology significantly enhances the capabilities of robotic teams in dangerous and low-visibility conditions.