In recent years, unmanned aerial vehicle (UAV) has been widely applied in various fields. Most detection is limited to fixed-view scenarios, while UAV-based object detection has a broader view and can monitor wider areas. However, the size of vehicles in images may change significantly and become smaller due to the changing and increasing distance between the drone and the ground, thus making vehicle detection challenging. In order to address the issue, a vehicle detection method for UAV aerial images based on ObjectBox was proposed. We replaced the original neck network with FPN to adapt to multi-scale object detection. Moreover, the Shuffle Attention module was added to FPN, improving its capability to obtain channel interaction information and spatial information by channel and spatial attention mechanisms. The experiments demonstrate the effectiveness of the improved ObjectBox detector. With a stronger neck network, we achieved a mean average accuracy of 55% and a real-time inference time of 39 FPS on the VisDrone sub-dataset. The proposed network keeps a balance between accuracy and speed to meet the requirements of practical applications.
Vehicle Detection Method in UAV Images Based on Improved ObjectBox
18.08.2023
4920373 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Vehicle detection method based on improved YOLOv3
IEEE | 2022
|Research on Improved YOLOv5 Vehicle Target Detection Algorithm in Aerial Images
DOAJ | 2024
|Vehicle Light Detection Method Based on Improved CenterNet
Springer Verlag | 2022
|Vehicle Detection in Aerial Images Based on YOLOv3
Springer Verlag | 2021
|