DV-YOLO: a deep learning framework for small-object detection in UAV-based remote sensing imagery with applications to smart logistics

Unmanned aerial vehicles (UAVs) are being increasingly adopted as flexible remote sensing platforms for smart logistics applications, including warehouse inventory, last-mile delivery supervision, traffic flow analysis, port operations, and infrastructure inspection. Despite their advantages, reliable object detection in UAV-based remote sensing imagery remains challenging due to small object sizes, dense object distributions, arbitrary orientations, and complex backgrounds commonly encountered in logistics environments. Although recent YOLO-based detectors have shown promising performance, their effectiveness is often limited in high-resolution aerial scenes and under practical computational constraints imposed by UAV platforms. To address these challenges, this paper proposes DV-YOLO, an enhanced deep learning framework tailored for object detection in UAV-based remote sensing imagery for logistics-oriented applications. The proposed model extends YOLOv9 through a deeper and wider backbone architecture coupled with optimized feature fusion strategies that jointly exploit spatial and semantic representations. A novel cross-path fusion network at deep feature map (CPFNDFM) is introduced to improve the detection of small and densely distributed logistics-related objects such as vehicles, containers, and infrastructure elements. In addition, a lightweight connection aggregation (CA) module, inspired by VoVNet and ShuffleNetV2, is integrated to enhance feature reuse while maintaining computational efficiency suitable for real-time UAV deployment. Furthermore, a challenging benchmark dataset, termed harder vision drone, is constructed by combining and refining samples from VisDrone and DOTA to better reflect real-world UAV remote sensing scenarios in logistics environments. Extensive experimental evaluations conducted on VisDrone 2021, DOTA v2, and the proposed dataset demonstrate that DV-YOLO consistently outperforms state-of-the-art detectors, achieving up to 3.5% improvement in mean average precision (mAP) compared with YOLOv9. These results highlight the potential of the proposed framework to support robust, accurate, and efficient aerial perception for smart logistics and UAV-based remote sensing applications.