Deep-Learning-Based Context-Aware Multi-Level Information Fusion Systems for Indoor Mobile Robots Safe Navigation.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous mobile cleaning robots require robust hazardous object detection and avoidance for safe operation.
• Conventional object detectors struggle with low-feature and occluded objects, leading to safety risks.
• Missed detections or false classifications of hazardous objects compromise robot operational safety.

Purpose of the Study:

• To develop a deep-learning-based framework for enhanced hazardous object detection and avoidance in autonomous mobile cleaning robots.
• To improve the detection confidence and accuracy of low-feature and occluded hazardous objects.
• To ensure safer robot navigation by implementing a reliable safe-distance-estimation function.

Main Methods:

• A context-aware multi-level information fusion framework was proposed.
• An image-level-contextual-encoding module was integrated with the Faster RCNN ResNet 50 object detector.
• A safe-distance-estimation function was developed using detection results and object depth data.
• The framework was trained using a custom dataset and fine-tuning techniques.

Main Results:

• The proposed framework demonstrated higher confidence in detecting low-featured and occluded hazardous objects compared to conventional methods.
• An average detection accuracy of 88.71% was achieved in real-time experiments.
• The system effectively improved the detection of challenging objects in indoor environments.

Conclusions:

• The deep-learning-based framework significantly enhances the safety of autonomous mobile cleaning robots.
• The context-aware multi-level information fusion approach effectively addresses limitations in detecting occluded and low-feature objects.
• The developed system provides a reliable solution for hazardous object detection and avoidance, crucial for autonomous robot navigation.