Autonomous Detection of Humans in Off-Limits Mountain Areas

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous systems require robust human detection, especially in challenging environments like off-limits mountains where human presence is rare.
• Traditional object detection algorithms (e.g., Convolution Neural Network-based) are computationally intensive and inefficient for continuous operation in low-occurrence event scenarios.

Purpose of the Study:

• To develop a time-efficient autonomous human detection system for environments with rare human occurrences.
• To reduce computational load while maintaining high accuracy in human detection.

Main Methods:

• A novel approach combining motion detection with object classification within a defined 'feasible human space'.
• Motion detection is performed periodically; object classification is triggered only when motion is detected within the feasible space.
• The system was compared against state-of-the-art algorithms like HOG detector, YOLOv7, and YOLOv7-tiny.

Main Results:

• The proposed system demonstrates comparable accuracy to existing state-of-the-art human detection algorithms.
• Achieved a significant improvement in computational speed, running 62 times faster than YOLOv7 in experiments with no humans present.
• Introduced the novel concept of 'feasible human space' for targeted detection.

Conclusions:

• The proposed human detection system offers a highly efficient solution for environments with rare human events.
• It provides a practical and computationally inexpensive method for real-time autonomous human detection.
• The 'feasible human space' concept is a key innovation for optimizing detection performance and speed.