Paying more attention on backgrounds: Background-centric attention for UAV detection.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Unmanned Aerial Vehicles (UAVs) are increasingly used in critical applications like military reconnaissance and traffic monitoring.
• Detecting small, fast-moving UAVs presents significant challenges due to their size, speed, and limited onboard computational power.
• Existing detection methods often struggle with these limitations, necessitating innovative approaches.

Purpose of the Study:

• To introduce a novel Background-centric Attention Module (BAM) for improved UAV detection.
• To address the challenges posed by small object size, high speeds, and limited resources in UAV surveillance.
• To develop a method that leverages background context for more robust UAV identification.

Main Methods:

• Developed a Background-centric Attention Module (BAM) that focuses on background features rather than solely on UAV visual characteristics.
• Integrated the BAM into existing mainstream UAV detection frameworks, specifically YOLOv5 and TphPlus.
• Conducted extensive experiments on challenging datasets, including the Naval Postgraduate School Drones (NPS) and Flying drones (FLDrones) datasets.

Main Results:

• The BAM significantly enhanced the detection accuracy of UAVs across different datasets and detectors.
• The module improved performance without a substantial increase in computational time, demonstrating efficiency.
• Experiments validated the effectiveness of utilizing background information for UAV detection.

Conclusions:

• Background information is a crucial, often overlooked, feature for accurate UAV detection.
• The proposed BAM offers a computationally efficient and effective method for improving UAV surveillance systems.
• This research provides a new direction for UAV detection, inspired by human visual perception.