Oriented Vehicle Detection in Aerial Images Based on YOLOv4

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Convolutional Neural Network (CNN)-based object detectors commonly use horizontal bounding boxes.
• Objects in aerial imagery, like vehicles and buildings, often have distinct orientations and can be densely packed.
• Existing oriented bounding box methods struggle with discontinuous boundaries due to angle periodicity or corner ordering.

Purpose of the Study:

• To develop a simple and efficient oriented object detector.
• To overcome the limitations of existing methods in handling object orientation and boundary discontinuities.
• To improve detection speed and accuracy for oriented objects in aerial images.

Main Methods:

• Proposed a novel oriented object detection method integrated into the YOLOv4 architecture.
• Regressed the offset of an object's front point instead of angle or corners to avoid boundary problems.
• Introduced an Intersection over Union (IoU) correction factor for stable training.

Main Results:

• Achieved significantly higher detection speed compared to other methods.
• Maintained high accuracy, outperforming existing approaches on DOTA and HRSC2016 datasets.
• Obtained the highest mean Average Precision (mAP) for classes with clear front-side appearances on the DOTA dataset.
• Demonstrated over 25% increase in detection speed due to the efficient YOLOv4 architecture.

Conclusions:

• The proposed front-point regression method effectively addresses boundary discontinuity issues in oriented object detection.
• The integration with YOLOv4 provides a highly efficient and accurate solution for aerial object detection.
• The method shows superior performance, particularly for objects with prominent front-side characteristics in aerial imagery.