Unmanned aerial vehicles for human detection and recognition using neural-network model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human action recognition is vital for machine understanding of behavior.
• Drone-based RGB videos present challenges like dynamic backgrounds, motion blur, and occlusions.
• Existing methods struggle with the complexities of aerial footage.

Purpose of the Study:

• To develop a robust method for human action recognition in challenging drone-recorded RGB videos.
• To improve the accuracy and reliability of action recognition systems in real-world scenarios.
• To address the limitations of current approaches in handling diverse human movements and video conditions.

Main Methods:

• Video frame segmentation and preprocessing to enhance foreground object visibility.
• Human body detection using the YOLOv9 algorithm.
• Skeleton extraction, feature engineering (positions, angles, distances), and 3D point cloud generation.
• Data optimization with Kernel Discriminant Analysis (KDA) and classification via a deep Convolutional Neural Network (CNN).

Main Results:

• The proposed model achieved action recognition accuracies of 0.68 on UAV-Human, 0.75 on UCF, and 0.83 on Drone-Action datasets.
• Effective handling of challenges such as motion blur, occlusions, and varying viewpoints.
• Demonstrated superior performance compared to existing methods on benchmark datasets.

Conclusions:

• The developed method offers a significant advancement in human action recognition from drone footage.
• The approach is effective in overcoming common challenges associated with aerial video analysis.
• This work paves the way for more sophisticated human-aware systems in surveillance, robotics, and entertainment.