Vehicle detection in UAV images is constrained by the differences in environmental and technical conditions such as target type and size, degree of occlusion, acquisition altitude, illumination, and weather. High levels of false alarms and missed detections are the most prominent challenges faced by traditional detection algorithms, particularly in the case of occluded or tiny targets acquired by UAV. This research sustains such problems by employing data augmentation and transfer learning techniques with an adjustable YOLOv9 model on UAV image dataset. We apply focused data augmentation techniques to enhance model performance over varied use cases, and transfer learning turns out to be effective in adapting the model to work for UAV-based occluded vehicle detection. The targeted improvements in the modified architecture of YOLOv9 enables effective detection of vehicles, including occluded ones with different sizes and orientations, thus, minimizing false alarms. The effectiveness of the proposed approach was experimentally validated, outperforming the existing ones and providing an architecturally and practically adequate solution for vehicle detection on real UAVs.
Vehicle Detection in UAV Images Using Data Augmentation and Transfer Learning with Modified YOLOv9
2024-12-02
511661 byte
Conference paper
Electronic Resource
English