Dynamic convolutional neural networks for altitude aware UAV object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Object detection on Unmanned Aerial Vehicles (UAVs) faces challenges from varying altitudes and limited onboard computational resources.
• Traditional Convolutional Neural Networks (CNNs) are often inefficient for real-time UAV deployment due to high computational and memory demands.

Purpose of the Study:

• To develop a resource-efficient and accurate object detection system for UAVs by adapting CNN configurations to altitude.
• To investigate the impact of altitude-specific parameter tuning (image resolution, network width, kernel size) on detection performance and efficiency.

Main Methods:

• Modeled UAV object detection as a context-aware task using altitude as an adaptive parameter.
• Trained specialized CNN configurations for discrete altitude ranges.
• Implemented a dynamic parameter switching mechanism based on altitude thresholds.

Main Results:

• Optimal CNN settings for object detection vary significantly across different altitudes.
• A dynamic CNN framework with altitude-based switching achieved significant resource savings with minimal accuracy loss compared to non-dynamic models.
• Threshold analysis demonstrated the trade-off between accuracy and efficiency.

Conclusions:

• Altitude-specific adaptation of CNNs is crucial for efficient UAV object detection.
• The proposed dynamic CNN framework provides a scalable and energy-efficient solution for various UAV applications.
• This approach enhances the practicality of real-time object detection in resource-constrained aerial platforms.