Vehicle Detection with Orientation from Drones: A Comparison of Oriented-Based vs. Rotated-Based Detection Methods

Tran Anh Dung, Tran Tue Tanh, Khanh-Duy Nguyen

Abstract: With the rapid development of Unmanned Aerial Vehicles (UAV s), object detection in aerial images has garnered increasing attention from both industry and academia. However, detecting objects in aerial images presents significant challenges due to their varying orientations. This work addresses these challenges by employing rotated and oriented object detection methods, which utilize rotated bounding boxes to detect objects. Our contributions are two-fold. First, we developed a dataset called UIT-Drones-OVD (Oriented Vehicle Detection) to evaluate orientation-based object detection methods. The dataset comprises 881 drone-captured aerial images and includes four categories: pedestrian, motorbike, car, and bus. Second, we conducted experiments on five state-of-the-art methods from two primary approaches: oriented-based methods and rotated-based methods. These methods are recognized for their effectiveness in detecting and localizing objects with various orientations, a critical capability in aerial imagery where objects often appear at arbitrary angles. Experimental results demonstrate that oriented-based methods consistently outperformed rotated-based methods in our tests, likely due to their ability to more accurately capture the complex orientations of objects in aerial images, while rotated-based methods rely on fixed bounding box orientations.

External IDs:dblp:conf/iccais/DungTN24