A Global-Local Self-Adaptive Network for Drone-View Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning has significantly advanced object detection, but drone-view object detection presents unique challenges.
• Tiny-scale and blurred objects in aerial imagery provide less information for accurate detection.
• Uneven object distribution, especially crowded regions, leads to inefficient detection in drone surveillance.

Purpose of the Study:

• To propose an end-to-end global-local self-adaptive network (GLSAN) for robust and efficient drone-view object detection.
• To enhance the detection of small, blurred, and crowded objects in aerial images.
• To improve the overall performance and adaptivity of object detection systems for drone applications.

Main Methods:

• Developed a global-local detection network (GLDN) integrating a global-local fusion strategy for precise detection.
• Introduced a self-adaptive region selecting algorithm (SARSA) for unsupervised, dynamic cropping of crowded image regions.
• Utilized a local super-resolution network (LSRN) to enlarge cropped images, enhancing feature extraction and aiding data augmentation.

Main Results:

• The GLSAN demonstrated effectiveness and adaptivity on benchmark datasets like VisDrone2019-DET and UAVDT.
• The proposed SARSA and LSRN components contributed to improved robustness and efficiency in object detection.
• The network showed proven advantages when applied to the DroneBolts dataset for industrial applications.

Conclusions:

• The GLSAN effectively addresses key challenges in drone-view object detection, offering improved accuracy and efficiency.
• The self-adaptive region selection and super-resolution components are crucial for handling complex aerial scenes.
• The method shows significant potential for real-world industrial applications requiring precise aerial object detection.