1. Introduction
In recent years, the rapid iteration and intelligent upgrading of UAV drone technology have brought unprecedented opportunities to numerous fields. With their flexible deployment, low cost, and high-altitude panoramic observation capabilities, UAV drones have been increasingly applied in areas such as intelligent traffic management, power line inspection, and oil and gas pipeline monitoring. Compared with traditional ground-level perspectives, the unique vantage point of UAV drones allows for efficient coverage of large areas and the acquisition of global information. For instance, in intelligent transportation, they can monitor traffic density and illegal lane changes in real time; in power grid inspection, they can quickly identify potential hazards on high-voltage lines. However, this elevated perspective also introduces significant imaging challenges. Due to the shooting altitude, targets such as vehicles, pedestrians, and tower markers appear extremely small in the images, with low pixel occupancy and blurred morphology. Moreover, in complex scenarios, these small targets often overlap with each other or are partially occluded by vegetation or buildings, leading to the loss of critical features. Therefore, the development of methods for detecting small objects in UAV drone aerial imagery is of great importance.
To address the specific challenges of dense small target detection in UAV drone images, researchers have proposed various approaches. Some methods focus on improving feature extraction backbones, others on enhancing multiscale fusion, and still others on designing specialized detection heads. However, existing models still suffer from limited global perception, degraded upsampling details, and bottlenecks in cross-layer feature fusion. These issues result in high rates of missed detections and false positives for small targets in complex environments. In response, we propose a novel model called SOD-YOLO (Small Object Detection-You Only Look Once), which integrates multi-scale global perception, content-aware upsampling, and cross-layer semantic enhancement.
Specifically, our contributions are as follows. First, we replace the original convolutional backbone of YOLOv8 with a Swin Transformer-Tiny, which enhances the model’s ability to model long-range dependencies and capture multi-scale contextual information. Second, we introduce the CARAFE (Content-Aware ReAssembly of Features) operator to replace the conventional upsampling method. This improves the recovery of fine details and the preservation of object boundaries during feature map upsampling. Third, we design a novel C2f_RVB module by integrating RepVITBlock, which innovates the feature fusion mechanism and boosts the efficiency of cross-layer semantic integration. Through these three improvements, our model effectively tackles the performance weaknesses of existing methods in complex UAV drone scenes.
2. Related Work
The task of object detection in UAV drone aerial images has attracted considerable attention. Many studies have attempted to adapt generic object detectors to the unique characteristics of aerial scenes. For example, some researchers have proposed an improved RT-DETR method for UAV drone images, which effectively reduces false and missed detections. Others have developed lightweight models based on YOLOv11 for edge deployment while maintaining satisfactory small target detection. Some works have explored Transformer-based detection heads with high-order spatial feature extraction for accurate small target localization. Additionally, multi-scale fusion and high-resolution feature enhancement have been employed to handle the significant scale variations in UAV drone imagery. An improved YOLOX algorithm has also been presented to address low confidence and missed detections caused by dense small targets in cluttered backgrounds.
Despite these advances, most existing models still face limitations. Convolutional backbones inherently lack global context, traditional upsampling methods often wash out fine details, and simple concatenation or addition-based fusion schemes cannot fully exploit complementary information from different semantic levels. Our proposed SOD-YOLO model addresses these gaps by synergistically combining a Transformer backbone, content-aware upsampling, and a strengthened cross-layer fusion block.
3. Proposed Method: SOD-YOLO
Our SOD-YOLO model is built upon the YOLOv8 architecture. The overall framework is shown in the conceptual diagram (see the image below). The input image first passes through the backbone network, which is a Swin Transformer-Tiny. After multi-stage feature extraction, we obtain three initial feature maps at different scales. These are then fed into the feature pyramid network, where we replace all standard upsampling layers with CARAFE modules and replace all original C2f modules with our designed C2f_RVB modules. The processed multi-scale features are then passed to the detection head, which predicts bounding boxes and class probabilities.
3.1 Swin Transformer-Tiny Backbone
In the original YOLOv8, the backbone is composed of convolutional layers. While efficient, convolution operations have a limited receptive field, making it difficult to capture long-range dependencies and global context. This is particularly problematic for UAV drone images, where small targets are scattered across a large spatial area and their relationships are crucial for accurate detection. We therefore adopt Swin Transformer-Tiny as the backbone.
The Swin Transformer-Tiny architecture consists of four stages. In Stage 1, the input image (H×W×3) is first processed by a Patch Partition layer and a Linear Embedding layer, resulting in feature maps of size H/4×W/4 with C channels. This is followed by two Swin Transformer Blocks. The Swin Transformer Block uses a window-based multi-head self-attention (W-MSA) mechanism and a shifted window multi-head self-attention (SW-MSA) mechanism alternately. This design allows local window attention with significantly reduced computational complexity while still enabling cross-window information exchange. Each block also incorporates layer normalization (LN) and a multi-layer perceptron (MLP). In subsequent stages, patch merging layers halve the spatial resolution and double the channel dimensions, producing multi-scale features at resolutions of H/8×W/8, H/16×W/16, and H/32×W/32. The number of Swin Transformer Blocks is 2, 2, 6, and 2 for the four stages respectively.
By using Swin Transformer-Tiny, our model gains the ability to model long-range dependencies through self-attention, while the hierarchical structure captures multi-scale information. This is especially beneficial for UAV drone imagery where objects of very different sizes coexist. The improved global perception helps reduce false positives by better understanding the scene context.
3.2 CARAFE Upsampling
In the feature pyramid network of YOLOv8, nearest-neighbor or bilinear interpolation followed by convolution is typically used for upsampling feature maps. These methods are content-agnostic and often produce blurry boundaries and lost details, which are detrimental for small target detection. To address this, we replace all upsampling layers with CARAFE (Content-Aware ReAssembly of Features).
CARAFE consists of two main modules: a kernel prediction module and a content-aware reassembly module. Given an input feature map of size H×W×C, the kernel prediction module first compresses the channel dimension via a 1×1 convolution to reduce computational cost. Then, a content encoding subnetwork applies a large kernel convolution (e.g., 5×5) to generate a kernel weight map of size H×W×(k_up^2×σ^2), where k_up is the upsampling kernel size and σ is the upsampling factor. The kernel weights are normalized using a softmax function across the local region. The content-aware reassembly module then uses these learned kernels to reassemble features from the input map, producing an output feature map of size σH×σW×C. This process adaptively emphasizes important spatial details, such as edges and corners of small objects.
The integration of CARAFE in our SOD-YOLO model significantly improves the quality of upsampled feature maps. Small objects that were previously blurred now have sharper boundaries, and the network can better distinguish them from background clutter. This leads to higher recall and precision for dense small targets in UAV drone images.
3.3 C2f_RVB Module
The original YOLOv8 uses the C2f module in its neck, which consists of a series of convolutions and shortcut connections. Although effective, the purely convolutional design limits the ability to capture long-range interactions across feature levels. To enhance cross-level semantic fusion, we propose the C2f_RVB module, which incorporates RepVITBlock.
RepVITBlock is designed to combine the strengths of convolution and Transformer-like attention. As shown in the schematic, it first applies a 3×3 depthwise convolution and a 1×1 depthwise convolution for multi-scale feature extraction. Then, a squeeze-and-excitation (SE) module recalibrates channel-wise feature responses. Finally, a 1×1 convolution projects the features. This design is lightweight but provides a receptive field larger than standard convolutions and introduces a form of content-adaptive weighting.
Within the C2f_RVB module, the input feature is first passed through a convolution layer to adjust channels. Then, the feature is split into two paths. The first path passes through a series of RepVITBlocks (typically 3 blocks) to perform enhanced cross-level feature transformation. The second path is a skip connection that retains the original features. The two paths are then concatenated and fused by a convolution, batch normalization, and SiLU activation. This dual-path design preserves both the raw information and the semantically enriched features, effectively resolving the bottleneck of cross-layer fusion. The C2f_RVB module is used in all the fusion stages of the feature pyramid, replacing every original C2f module.
4. Experiments
4.1 Dataset
We evaluate our method on the VisDrone2019 dataset, which is a benchmark for object detection in UAV drone images. The dataset contains 10,209 static images captured from 14 different cities in China, covering diverse urban scenes, traffic hubs, and other environments. It includes annotations for 10 object categories: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor. The dataset is split into 6,471 training images, 548 validation images, and 3,190 test images. This dataset is challenging because objects are often very small, densely packed, and heavily occluded.
4.2 Experimental Setup
All experiments are conducted on a system with an Intel Core i9-13900F CPU, an NVIDIA GeForce RTX 4070 GPU with 12 GB memory, and Windows 11. The deep learning framework is PyTorch 2.0.1 with Python 3.8.5. We use the SGD optimizer with an initial learning rate of 0.001, batch size of 16, and train for 200 epochs. Mosaic data augmentation is employed throughout training to improve generalization. The loss curves show that both training and validation losses decrease rapidly in the first 50 epochs and then gradually converge, without any noticeable overfitting.
4.3 Evaluation Metrics
We adopt four standard metrics: Precision (P), Recall (R), F1-score (F), and mean Average Precision (mAP). Their definitions are as follows:
$$ P = \frac{TP}{TP + FP} $$
$$ R = \frac{TP}{TP + FN} $$
$$ F = 2 \cdot \frac{P \cdot R}{P + R} $$
$$ mAP = \frac{1}{C} \sum_{c=1}^{C} AP_c $$
where TP, FP, and FN are true positives, false positives, and false negatives, respectively. AP_c is the average precision for class c, and C=10 for the VisDrone dataset.
4.4 Ablation Study
To validate the contribution of each proposed improvement, we conduct ablation experiments. The baseline is the original YOLOv8 model. We then sequentially add Swin Transformer-Tiny, CARAFE, and C2f_RVB. The results are summarized in Table 1.
| Method | Swin-T | CARAFE | C2f_RVB | P (%) | R (%) | F (%) | mAP (%) |
|---|---|---|---|---|---|---|---|
| Baseline (YOLOv8) | × | × | × | 46.2 | 34.8 | 39.6 | 34.4 |
| Improvement 1 | √ | × | × | 49.3 | 39.3 | 43.7 | 39.0 |
| Improvement 2 | √ | √ | × | 50.2 | 39.3 | 44.0 | 39.3 |
| SOD-YOLO | √ | √ | √ | 50.4 | 39.7 | 44.4 | 39.6 |
From Table 1, we observe the following:
- Replacing the backbone with Swin Transformer-Tiny (Improvement 1) yields a significant gain of 4.6% in mAP and 4.1% in F1-score. This demonstrates that the Transformer backbone dramatically improves global context modeling, enabling better detection of small targets distributed across the image.
- Adding CARAFE (Improvement 2) further improves precision by 0.9% and F1-score by 0.3%, while mAP increases by 0.3%. The content-aware upsampling effectively preserves fine details, leading to sharper boundaries and fewer false positives.
- Finally, incorporating C2f_RVB (full SOD-YOLO) raises recall by 0.4% and mAP by 0.3%. The enhanced cross-layer fusion better integrates semantic and spatial information, which is crucial for detecting tightly packed small objects.
All three improvements contribute positively, and the final SOD-YOLO model achieves the best overall performance.
4.5 Comparison with State-of-the-Art Methods
We compare our SOD-YOLO with several popular object detectors, including YOLOv5, YOLOv8, YOLOv10, Fast RCNN, SSD, and a recent small-object-specific model DEPA-YOLO. All models are trained and tested on the same VisDrone2019 dataset under the same settings. The quantitative results are shown in Table 2.
| Model | P (%) | R (%) | F (%) | mAP (%) |
|---|---|---|---|---|
| YOLOv5 | 45.1 | 34.6 | 39.1 | 33.9 |
| YOLOv8 | 46.2 | 34.8 | 39.6 | 34.4 |
| YOLOv10 | 46.4 | 35.0 | 39.9 | 35.0 |
| Fast RCNN | 43.1 | 37.6 | 40.1 | 33.7 |
| SSD | 32.5 | 24.1 | 27.6 | 30.9 |
| DEPA-YOLO | 49.5 | 37.2 | 42.4 | 38.2 |
| SOD-YOLO (Ours) | 50.4 | 39.7 | 44.4 | 39.6 |
Our SOD-YOLO surpasses all competing methods across all metrics. Specifically, it achieves the highest mAP of 39.6%, which is 4.6% higher than YOLOv10, 5.2% higher than YOLOv8, and 1.4% higher than DEPA-YOLO. The F1-score of 44.4% also outperforms the second-best (DEPA-YOLO) by 2.0%. The precision and recall are both the highest, indicating that our model correctly identifies more small targets while maintaining fewer false detections. This demonstrates the effectiveness of our integrated improvements for dense small object detection in UAV drone imagery.
4.6 Qualitative Results
Visual comparisons further illustrate the superiority of SOD-YOLO. In a typical UAV drone scene with many overlapping small cars and pedestrians, the baseline YOLOv8 misses several small vehicles and produces false positives due to background clutter. YOLOv10 and Fast RCNN also show missed detections and inaccurate bounding boxes. In contrast, SOD-YOLO consistently detects almost all visible small targets, with tightly fitting bounding boxes and high confidence scores. The improvements are especially noticeable for targets that are partially occluded or located near image boundaries.
5. Urban Low-Altitude Traffic Dense Small Target Detection Platform
To bridge the gap between algorithmic research and practical application, we have developed a detection platform based on the SOD-YOLO model. This platform is implemented in Python 3.8 with a PyQt5 graphical user interface. It is designed specifically for urban low-altitude traffic scenarios, where dense small targets such as pedestrians, bicycles, and vehicles need to be monitored in real time.
The platform supports three main functions:
- Image batch detection: Users can load a folder of UAV drone images and perform batch detection. The results are saved with bounding boxes and class labels.
- Video offline detection: Recorded video files can be processed frame by frame, with detection results overlaid on the video and stored for later analysis.
- Real-time camera detection: By connecting an external camera (e.g., a USB camera or a video stream from a UAV drone’s onboard camera), the platform can perform live detection and display results instantly.
The interface includes buttons for importing files, starting detection, saving results, and adjusting display settings. The platform is designed to be compatible with various hardware, including PCs and embedded devices, making it suitable for deployment in control rooms or on mobile inspection vehicles. This practical tool effectively addresses the inefficiencies and high missed-detection rates of manual inspection in urban low-altitude traffic management.
6. Conclusion
In this work, we addressed the challenging problem of dense small target detection in UAV drone aerial images. We proposed SOD-YOLO, a model that integrates multi-scale global perception, content-aware upsampling, and cross-layer semantic enhancement. Specifically, we replaced the convolutional backbone with Swin Transformer-Tiny to better capture long-range dependencies, introduced CARAFE to improve upsampling detail restoration, and designed a C2f_RVB module to enhance cross-level feature fusion. Extensive experiments on the VisDrone2019 dataset validated the effectiveness of each component. Our final model achieved a mAP of 39.6% and an F1-score of 44.4%, significantly outperforming state-of-the-art detectors such as YOLOv8, YOLOv10, and DEPA-YOLO. Furthermore, we developed a practical detection platform for urban low-altitude traffic scenarios, demonstrating the real-world applicability of our approach. Future work could explore further improvements in model efficiency for edge deployment and extend the framework to handle multi-modal inputs (e.g., thermal or infrared images) for enhanced robustness in various weather and lighting conditions. Overall, SOD-YOLO provides a reliable solution for precise localization and recognition of dense small objects in UAV drone imagery.
References
[1] Wu Yiquan, Tong Kang. Research Progress on Small Object Detection in UAV Aerial Images Based on Deep Learning. Acta Aeronautica et Astronautica Sinica, 2025, 46(3): 181-207.
[2] Hu Yanglin, Zhang Tiankui, Li Bo, et al. A Survey on UAV-Enabled Integrated Sensing and Communication Networking and Technologies. Journal of Electronics & Information Technology, 2025, 47(4): 859-875.
[3] Yuan Yubin, Wu Yiquan, Zhao Langyue, et al. Research Progress on Multi-Object Detection and Tracking in UAV Aerial Videos Based on Deep Learning. Acta Aeronautica et Astronautica Sinica, 2023, 44(18): 6-36.
[4] Wu Han, Sun Hao, Liu Kui, et al. Research Progress on Multi-Object Feature Association in UAV Videos. Acta Aeronautica et Astronautica Sinica, 2026, 47(4): 131-161.
[5] Zeng Fanshuang, Zhang Naiwen, Li Yingzhuo, et al. Ground Small Target Detection System Based on YOLOv8 Model and Streamlit. Modern Information Technology, 2025, 9(19): 57-61,69.
[6] Zeng Jing, Liao Shuzhen, Zhou Yongfu, et al. Helmet Detection and Recognition System Based on YOLOv10+PyQt5. Modern Information Technology, 2025, 9(19): 22-25.
[7] Cheng Jiapei, Wang Xusheng, Lin Qibin, et al. Improved YOLOv5 for Multi-Object Recognition and Tracking of Pedestrians and Vehicles. Modern Information Technology, 2025, 9(17): 73-77,82.
[8] Li Pan, Lou Li. Hybrid Wafer Map Defect Pattern Recognition Based on Vision Transformer. Modern Information Technology, 2025, 9(19): 26-30.
[9] Zhong Shuai, Wang Liping. MCS-RETR: Improved RT-DETR for Object Detection in UAV Aerial Images. Acta Aeronautica et Astronautica Sinica, 2025, 46(22): 260-276.
[10] Tu Yuzhi, Wang Faxiang, Wu Chunlin. Lightweight UAV Aerial Small Target Detection Model Integrating Multi-Attention Mechanisms. Computer Engineering and Applications, 2025, 61(11): 93-104.
[11] Zhang Xuanyu, Zhou Sihang, Huang Jian, et al. Small Object Detection Algorithm for UAV Aerial Images Based on High-Order Spatial Feature Extraction. Computer Engineering and Applications, 2025, 61(12): 210-221.
[12] Chen Zhiwang, Xiao Dichuang, Lyu Changhao, et al. UAV Aerial Object Detection Based on Multi-scale Fusion and High-Resolution Feature Enhancement. Control and Decision, 2025, 40(7): 2290-2299.
[13] Zheng Zhanji, Feng Changkui, Zhao Yangyang, et al. Detection Method for Non-Motor Vehicle Small Objects in Dense Scenes from UAV Aerial Perspective. Journal of Transportation Systems Engineering and Information Technology, 2025, 25(4): 147-161.
[14] Varghese R, Sambath M. YOLOv8: A Novel Object Detection Algorithm with Enhanced Performance and Robustness. 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), 2024: 1-6.
[15] Zhu P F, Wen L Y, Du D W, et al. Detection and Tracking Meet Drones Challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(11): 7380-7399.
[16] Liu Chenjie, Liu Wei, Yang Wendi, et al. DEPA-YOLO: A Small Object Detection Model from UAV Perspective. Journal of Computer Science and Exploration, 2026, 20(1): 266-279.