An Improved FBPN-Based Detection Network for Vehicles in Aerial Images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Deep learning has advanced object detection, but vehicle detection in aerial images remains challenging due to low resolution, complex backgrounds, and object rotation.
• Existing methods struggle with the unique difficulties presented by aerial imagery for vehicle identification.

Purpose of the Study:

• To propose an improved feature-balanced pyramid network (FBPN) for enhanced small object detection capabilities.
• To develop a robust vehicle detection framework for aerial images by integrating FBPN with a modified faster region convolutional neural network (faster-RCNN).

Main Methods:

• An improved feature-balanced pyramid network (FBPN) was developed to boost small object detection.
• The FBPN was integrated with a modified faster region convolutional neural network (faster-RCNN) to create a specialized aerial vehicle detection framework.
• Focal loss was employed to mitigate the class imbalance between easy and hard detection samples.

Main Results:

• The proposed framework demonstrated superior performance in vehicle detection on aerial images.
• Experimental results on VEDIA, USCAS-AOD, and DOTA datasets confirmed the framework's effectiveness.
• The integration of FBPN and faster-RCNN significantly improved detection accuracy for small and complexly oriented vehicles.

Conclusions:

• The developed framework offers a significant advancement in aerial vehicle detection.
• The improved FBPN and focal loss contribute to overcoming the limitations of detecting vehicles in challenging aerial imagery.
• This research provides a more effective solution for automated vehicle surveillance and analysis from aerial platforms.