Accurate real-time object detection in unmanned aerial vehicle (UAV) is essential for surveillance, autonomous navigation, infrastructure inspection, and intelligent aerial monitoring. However, UAV imagery has significant challenges due to small object sizes, viewpoint variations, illumination changes, occlusions, and dynamic backgrounds. To address these limitations, this study proposes a hybrid UAV object detection framework that integrates sequential Generative Adversarial Network (GAN)-based dataset augmentation with YOLO-CornerNet detection framework. A new multimodal benchmark dataset, UAVision-M3 (Multimodal, Multi-scenario, Multi-object) was developed by combining UAV videos and high-resolution aerial imagery. To enhance the dataset diversity, a sequential CycleGAN-Conditional GAN (cGAN) augmentation pipeline was performed to adapt domain and condition-aware synthesis with annotation consistency. The proposed framework differs from traditional techniques, because a systematic comparative evaluation of YOLOv3-YOLOv9 was conducted under the same training environment to discover the most suitable backbone for hybrid integration. Experimental findings reveal that YOLOv8 achieves the optimal trade-off and is combined with CornerNet to improve localization of small and dense aerial objects through keypoint-based refinement. For UAVision-M3, the hybrid model achieves 95.1% Precision, 93.3% Recall and 97.8% mAP@0.5, which outperforms all YOLO versions alone. Extensive assessments on the VisDrone and UAVDT benchmarks further show persistent gains in detection accuracy and localization performance under tough airborne conditions while preserving real-time inference capacity. These findings suggest that sequential GAN-based dataset augmentation and hybrid object localization is a robust and successful approach for UAV object detection in demanding aerial scenarios.