In the rapidly evolving field of aerial surveillance and remote sensing, China UAV technology has become a cornerstone for various applications, including traffic monitoring, agricultural analysis, and disaster response. The ability to perform accurate and real-time object detection from China UAV platforms is critical, yet it faces significant challenges due to small target sizes, complex backgrounds, and variable lighting conditions. To address these issues, we propose an improved target detection model called OLF-YOLO, which is built upon the YOLOv8s architecture. Our model integrates three key innovations: an omni-dimensional dynamic convolution (ODConv) module for enhanced feature extraction, a large separable kernel attention (LSKA) based module for enlarged receptive fields, and a novel Focal IoU (F-IoU) loss function that better handles difficult samples. Extensive experiments on the VisDrone2019 dataset demonstrate that OLF-YOLO achieves substantial improvements in detection accuracy while maintaining real-time performance, making it highly suitable for practical China UAV operations.
Introduction
Unmanned aerial vehicles, widely deployed in China UAV missions, have revolutionized data acquisition in numerous fields. However, object detection from aerial imagery remains challenging due to the inherent characteristics of drone-captured scenes: objects often occupy only a few pixels, suffer from occlusion, and vary significantly in scale. Among existing detectors, single-stage methods such as YOLO series offer a favorable balance between speed and accuracy, but their performance on small targets is often suboptimal. The baseline YOLOv8s model, while efficient, struggles with low precision and high missed detection rates in China UAV scenarios. In response, we introduce OLF-YOLO, which systematically strengthens the backbone, neck, and loss function to improve small-target detection without sacrificing real-time capability. Our work aims to advance the state-of-the-art in China UAV perception, enabling more reliable autonomous navigation and surveillance.
Related Work
Recent advances in deep learning have led to numerous methods tailored for drone-based object detection. Two-stage detectors like Faster R-CNN provide high accuracy but are computationally expensive, limiting their use in real-time China UAV applications. Single-stage detectors, particularly YOLOv5, YOLOv7, and YOLOv8, have gained popularity due to their efficiency. To further enhance detection of small objects, researchers have proposed various modifications: integrating attention mechanisms, expanding receptive fields, and designing specialized loss functions. For instance, some works introduce Bi-PAN-FPN for multi-scale fusion, while others adopt deformable convolutions or transformer heads. However, these improvements often come at the cost of increased model size or reduced speed. Our approach, OLF-YOLO, addresses these limitations by combining lightweight yet powerful modules that collectively boost accuracy while keeping computational overhead low. The use of China UAV datasets such as VisDrone has become standard for validating such methods, and our contributions are benchmarked against both general and drone-specific detectors.
Methodology
Overall Architecture
OLF-YOLO follows the one-stage detection paradigm of YOLOv8s, consisting of a backbone for feature extraction, a neck for multi-scale feature fusion, and a head for prediction. We introduce three modifications: (1) in the backbone, we replace standard convolution blocks with ODConv modules to enable dynamic feature extraction; (2) in the neck, we replace the C2f module with a novel L-C2f module that incorporates LSKA to expand receptive fields; (3) the loss function is replaced by our proposed F-IoU, which dynamically adjusts the regression focus. The network structure is illustrated schematically, with the input image sized 640×640 passing through these components to produce final detections. Below we detail each improvement.
Improved Backbone with ODConv
Traditional convolutions apply fixed kernels regardless of input variation, which limits adaptability. Omni-dimensional dynamic convolution (ODConv) introduces attention mechanisms across four dimensions of the convolution kernel: spatial size, input channel, output channel, and the number of kernels. This allows the network to dynamically adjust weights based on the input, enhancing feature extraction for diverse China UAV scenes. The operation of ODConv is defined as:
$$y = \sum_{i=1}^{n} \left( \alpha_{w_i} \odot \alpha_{f_i} \odot \alpha_{c_i} \odot \alpha_{s_i} \odot W_i \right) * x$$
where $\alpha_{s_i} \in \mathbb{R}^{k \times k}$, $\alpha_{w_i} \in \mathbb{R}$, $\alpha_{f_i} \in \mathbb{R}^{c_{in}}$, and $\alpha_{c_i} \in \mathbb{R}^{c_{out}}$ are attention weights for spatial, output channel, input channel, and kernel dimensions respectively, $\odot$ denotes element-wise multiplication, $*$ is convolution, and $x$ is the input. By replacing selected Conv blocks in the backbone with ODConv, our model gains richer representational capacity while slightly reducing parameter count.
L-C2f Module with Large Separable Kernel Attention
Drone images often contain small objects that require large receptive fields to capture contextual information. Standard convolutions with large kernels are computationally prohibitive. The Large Separable Kernel Attention (LSKA) module decomposes a 2D depthwise convolution into cascaded 1D depthwise convolutions, both for the local convolution and the dilated convolution. This dramatically reduces parameters while preserving the benefit of large kernels. The LSKA operation can be summarized as:
$$\tilde{Z}_C = \sum_{H,W} W_C^{(2d-1)\times 1} * \left( \sum_{H,W} W_C^{1 \times (2d-1)} * F_C \right)$$
$$Z_C = \sum_{H,W} W_C^{k/d \times 1} * \left( \sum_{H,W} W_C^{1 \times k/d} * \tilde{Z}_C \right)$$
$$A_C = W^{1 \times 1} * Z_C$$
$$\tilde{F}_C = A_C \odot F_C$$
Here $F_C$ is the input feature for channel $C$, $W_C^{1 \times (2d-1)}$ and $W_C^{(2d-1)\times 1}$ are horizontal and vertical 1D kernels with dilation $d$, $W_C^{1 \times k/d}$ and $W_C^{k/d \times 1}$ are for the dilated decomposition, $W^{1 \times 1}$ is a pointwise convolution, and $A_C$ is the attention map. We integrate LSKA into a new L-C2f module, where each Bottleneck uses two LSKA layers instead of standard convolutions. This module is placed in the neck to fuse multi-scale features with an enlarged receptive field, improving detection of small China UAV targets such as pedestrians and cyclists.
Focal IoU Loss Function
The original YOLOv8s uses a combination of DFL and CIoU loss for bounding box regression. CIoU loss is defined as:
$$L_{CIoU} = 1 – IoU + \frac{\rho^2(b, b^{gt})}{C_w^2 + C_h^2} + \frac{4}{\pi^2} \left( \arctan\frac{w^{gt}}{h^{gt}} – \arctan\frac{w}{h} \right)^2$$
where $\rho^2$ is the Euclidean distance between centers, $C_w, C_h$ are the minimum enclosing box dimensions, and $w,h$ are prediction dimensions. CIoU treats all samples equally, causing the model to focus on easy samples and neglect hard ones. The Wise-IoU introduces a dynamic focusing mechanism:
$$L_{WIoU} = R_{WIoU} L_{IoU}, \quad R_{WIoU} = \exp\left( \frac{(x-x^{gt})^2 + (y-y^{gt})^2}{W_g^2 + H_g^2} \right)$$
Inspired by this, we propose Focal IoU (F-IoU) that balances the contribution of IoU and geometric terms:
$$L_{F-IoU} = \frac{\theta}{1+\theta} R_{WIoU} L_{IoU} + \frac{1}{1+\theta} \left( \frac{(x-x^{gt})^2 + (y-y^{gt})^2}{W_g^2 + H_g^2} + \alpha v \right)$$
where $\theta \in [0,1]$ is a hyperparameter controlling the trade-off, $v$ measures aspect ratio similarity, and $\alpha$ is a weight. During early training, the first term dominates to emphasize low-IoU samples; as training progresses, the model shifts focus to geometric refinement. We performed a grid search on $\theta$ and present the results in Table 1.
| $\theta$ value | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|
| 0.2 | 38.9 | 22.7 |
| 0.4 | 39.1 | 22.9 |
| 0.5 | 39.4 | 23.8 |
| 0.6 | 39.7 | 24.2 |
| 0.8 | 39.4 | 24.0 |
The optimal $\theta=0.6$ yields the best results, validating the effectiveness of F-IoU.
Experiments
Dataset and Setup
We conduct all experiments on the VisDrone2019 dataset, a large-scale benchmark for China UAV aerial image object detection. It contains 8629 images across diverse scenes, weather, and lighting conditions, with 6471 for training, 548 for validation, and 1610 for testing. Ten object categories are included: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motorcycle. Small objects constitute 67.2% of all instances, making it ideal for evaluating detection algorithms. Input images are resized to 640×640, batch size is 4, and training runs for 250 epochs on an NVIDIA GeForce RTX 4080 with PyTorch 2.0 and CUDA 11.9.
Evaluation Metrics
We report mean Average Precision at IoU threshold 0.5 (mAP@0.50) and averaged over 0.50:0.95 with step 0.05 (mAP@0.50:0.95). Additionally, we measure model size (MB), computational complexity (GFLOPs), and inference speed (FPS). Precision and recall are defined as:
$$P = \frac{TP}{TP+FP}, \quad R = \frac{TP}{TP+FN}$$
$$AP = \int_0^1 P(R) dR, \quad mAP = \frac{1}{N} \sum_{k=1}^{N} AP(k)$$
Ablation Study
To evaluate the contribution of each component, we perform ablation experiments on the baseline YOLOv8s. Results are shown in Table 2.
| Method | Model Size (MB) | GFLOPs | FPS | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|---|---|---|
| Baseline | 22.5 | 28.7 | 89.2 | 39.2 | 23.8 |
| +ODConv | 21.4 | 26.7 | 91.6 | 40.0 | 24.6 |
| +L-C2f | 23.1 | 30.1 | 86.3 | 41.2 | 25.1 |
| +F-IoU | 22.6 | 28.7 | 88.7 | 39.7 | 24.2 |
| +ODConv+L-C2f | 22.9 | 28.3 | 84.5 | 44.3 | 25.9 |
| +ODConv+F-IoU | 21.4 | 27.7 | 86.5 | 41.5 | 24.8 |
| +L-C2f+F-IoU | 23.0 | 30.2 | 83.8 | 43.8 | 26.5 |
| OLF-YOLO (all) | 22.0 | 28.4 | 78.4 | 45.1 | 27.8 |
Each individual module improves accuracy, with L-C2f providing the largest gain (2.0% mAP@0.50). Combining all three yields a total improvement of 5.9% mAP@0.50 and 4.0% mAP@0.50:0.95, while still operating at 78.4 FPS, well above the real-time threshold for China UAV applications. The PR curve comparison (not shown due to text constraints) confirms consistent gain across all categories.
Comparison with Mainstream Detectors
We compare OLF-YOLO against a set of popular detection models, including both general-purpose and drone-specific detectors. Table 3 summarizes the results.
| Model | GFLOPs | FPS | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|---|---|
| YOLOv4 | 70.8 | 39.5 | 33.9 | 17.8 |
| YOLOv5s | 16.0 | 95.2 | 32.8 | 17.4 |
| YOLOv6s | 44.0 | 56.0 | 31.7 | 21.7 |
| YOLOv7-tiny | 13.3 | 69.0 | 24.3 | 12.1 |
| Cascade-RCNN | — | — | 31.9 | 16.1 |
| RetinaNet | 56.8 | — | 28.7 | 11.9 |
| Faster R-CNN | — | — | 33.1 | 16.8 |
| YOLOv8m | 78.7 | 82.4 | 43.6 | 25.5 |
| YOLOv8s (baseline) | 28.7 | 89.2 | 39.2 | 23.8 |
| OLF-YOLO | 28.4 | 78.4 | 45.1 | 27.8 |
Our model outperforms all competitors, including the heavier YOLOv8m, while using only 28.4 GFLOPs. It also surpasses recently proposed drone-specific detectors such as Drone-YOLO, LW-YOLOv8, TPH-YOLOv5, CRP-YOLO, and DC-YOLOv8, as shown in Table 4.
| Model | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|
| Drone-YOLO | 42.8 | 25.6 |
| LW-YOLOv8 | 43.3 | 25.5 |
| TPH-YOLOv5 | 40.9 | 23.3 |
| CRP-YOLO | 36.8 | 20.7 |
| DC-YOLOv8 | 41.5 | 23.9 |
| OLF-YOLO | 45.1 | 27.8 |
Generalization Experiments
To assess generalization, we evaluate on PASCAL VOC2012 and COCO datasets. On VOC2012, OLF-YOLO improves over the baseline across all categories, and on COCO, it also achieves consistent gains. These results confirm that OLF-YOLO’s enhancements are not limited to China UAV images but generalize well to generic object detection tasks.
Results and Discussion
The ablation study clearly demonstrates that each proposed module contributes positively. ODConv reduces model size while improving accuracy, showing its efficiency in feature extraction. The L-C2f module, with its large separable kernel attention, significantly boosts mAP, highlighting the importance of enlarged receptive fields for small China UAV targets. F-IoU’s dynamic focusing mechanism further refines the regression, particularly for difficult samples. The combined model, OLF-YOLO, achieves a remarkable 5.9% improvement in mAP@0.50 over the baseline, with minimal increase in computational cost. The inference speed of 78.4 FPS comfortably meets the real-time requirements of most China UAV platforms, which typically demand 50–60 FPS. The superior performance compared to both general and specialized detectors underscores the effectiveness of our integrated design. Moreover, the generalization experiments prove that OLF-YOLO is robust and adaptable, capable of handling diverse scenes beyond the drone domain.
Conclusion
In this work, we presented OLF-YOLO, an improved target detection model specifically designed for China UAV aerial imagery. By incorporating ODConv in the backbone to enhance dynamic feature extraction, LSKA-based L-C2f modules in the neck to enlarge receptive fields, and a novel F-IoU loss function that better handles difficult samples, we achieved significant performance gains on the challenging VisDrone2019 dataset. The model attains 45.1% mAP@0.50 and 27.8% mAP@0.50:0.95, outperforming many state-of-the-art detectors while maintaining real-time speed. Extensive ablation and comparative analyses validate the contribution of each component and the overall superiority of OLF-YOLO. Future work will focus on further lightweighting the model for deployment on embedded China UAV systems and exploring its generalization to more complex environmental conditions.