Surface and underwater human pose recognition based on temporal 3D point cloud deep learning.

Area of Science:

• Robotics and Computer Vision
• Remote Sensing Technologies
• Human-Computer Interaction

Background:

• Airborne optical cameras face limitations in simultaneously imaging surface and underwater environments due to visible-light wavelength constraints.
• Accurate human pose recognition is vital for safety and surveillance, particularly in detecting individuals in distress or drowning situations.

Purpose of the Study:

• To propose and validate a novel method using light detection and ranging (LiDAR) for simultaneous surface and underwater human pose recognition.
• To develop a neural network capable of recognizing human poses from irregular, temporal point-cloud data.

Main Methods:

• Construction of a temporal point-cloud dataset for surface and underwater human pose recognition.
• Application of radius outlier removal (ROR) and statistical outlier removal (SOR) for data noise reduction.
• Optimization of recognition accuracy using PointNet++ with various secondary sampling methods and sample sizes.

Main Results:

• The proposed method successfully achieved simultaneous detection and pose recognition of humans on the surface and underwater.
• The highest recognition accuracy reached 97.5012% after optimizing sampling techniques and sizes.
• Demonstrated the effectiveness of the neural network designed for irregular data in pose recognition.

Conclusions:

• Light detection and ranging (LiDAR) offers a viable solution for overcoming the limitations of optical cameras in dual-environment human pose recognition.
• The developed neural network approach, combined with data preprocessing and optimization, significantly enhances the accuracy of human pose detection in complex scenarios.
• This research provides a robust foundation for advanced safety and surveillance systems in aquatic and surface environments.