UAV-YOLOv8: A Small-Object-Detection Model Based on Improved YOLOv8 for UAV Aerial Photography Scenarios

Gang Wang, Yanfei Chen, Pei An, Hanyu Hong, Jinghu Hu, Tiange Huang2023 · DOI

TL;DR

This paper introduces UAV-YOLOv8, an enhanced YOLOv8 model specifically designed for detecting small objects in UAV aerial imagery. By incorporating novel loss functions, attention mechanisms, and a multi-scale feature fusion network, the model significantly improves detection accuracy for small and challenging objects while maintaining efficiency.

What Problem Does This Paper Address?

Unmanned Aerial Vehicles (UAVs) are increasingly used for applications like infrastructure inspection and traffic monitoring. However, object detection in UAV imagery presents unique challenges: a high proportion of small objects, complex backgrounds, and limited onboard computational resources. Existing object detection models often struggle to achieve high accuracy for small objects without demanding excessive computational power, creating a trade-off between performance and efficiency that is critical for UAV deployment. This paper aims to bridge this gap by developing a more accurate and efficient object detection model tailored for these specific UAV scenarios.

Methods

The authors build upon the YOLOv8 architecture, a state-of-the-art one-stage object detector, and introduce several key improvements:

  • WIoU v3 Loss Function: They replace the standard bounding box regression loss with WIoU v3. This novel loss function employs a dynamic, non-monotonic mechanism to allocate gradient gains more effectively, allowing the model to focus on samples of ordinary quality and thus improving localization accuracy and generalization.
  • BiFormer Attention Mechanism: To enhance the model's ability to focus on critical information and suppress background noise, they integrate the BiFormer attention mechanism into the backbone network. BiFormer efficiently filters out irrelevant features, allowing the model to attend more precisely to important object details.
  • Focal FasterNet Block (FFNB) and Multi-Scale Feature Fusion: A new feature processing module, FFNB, is proposed, based on the FasterNet block. This module is designed for efficiency and improved receptive field. Crucially, the authors extend YOLOv8's original three detection scales to five by incorporating FFNB. This creates a multiscale feature fusion network that effectively integrates shallow (location-rich) and deep (semantic-rich) features, significantly reducing missed detections of small objects.

Datasets

The primary dataset used for training and evaluation is VisDrone2019. This dataset is a mainstream benchmark for UAV aerial photography, featuring diverse scenarios, object types, and challenging conditions like dense object distributions and small object sizes. The paper also mentions experiments on self-made data collected from the campus of Wuhan Institute of Technology and urban areas in Hubei Province, China, to further validate the model's generalizability.

Key Insights and Ideas

  • Dynamic Loss Functions Matter: The success of WIoU v3 highlights that simply improving IoU metrics isn't enough; how the model learns from different quality samples (easy vs. hard, high-quality vs. low-quality) is crucial for robust localization.
  • Attention for Efficiency and Focus: BiFormer demonstrates that attention mechanisms can be made computationally efficient while still significantly improving a model's ability to discern relevant features from complex backgrounds, a key challenge in UAV imagery.
  • Multi-Scale is Key for Small Objects: Extending the detection scales, especially by fusing shallow features, is a direct and effective strategy for tackling the pervasive problem of small object detection in UAV data. The FFNB module enables this without excessive computational overhead.

Key Results

The proposed UAV-YOLOv8 model demonstrates significant improvements over the baseline YOLOv8s:

  • Accuracy: Achieves a 7.7% increase in mean detection accuracy (mAP0.5) compared to the baseline YOLOv8s.
  • Small Object Performance: Shows substantial improvements in AP values for categories like pedestrians, people, and motors, often exceeding 10% gains.
  • Efficiency: While adding detection layers increases complexity, the model's overall parameter count and size are managed effectively, and it still aims for real-time detection. The BiFormer block, in particular, reduces model size and parameters.
  • Comparison: Outperforms other mainstream YOLO series models (YOLOv3, v4, v5, v7) and classical detection models (Faster R-CNN, RetinaNet, CenterNet, etc.) in terms of detection accuracy, especially for small objects.

Limitations and Future Work

The authors acknowledge that while the improved model is more complex and has a larger feature map size in shallow layers, leading to increased computation and inference time, there is still room for optimization in computational resource consumption. They also note that the detection accuracy for very small objects like bicycles remains a challenge, and they plan to focus on further optimizing detection accuracy for these objects while keeping resource constraints in mind for future research.

Summary updated

BibTeX
@misc{wang2023uav,
  author = {Gang Wang and Yanfei Chen and Pei An and Hanyu Hong and Jinghu Hu and Tiange Huang},
  title = {{UAV-YOLOv8: A Small-Object-Detection Model Based on Improved YOLOv8 for UAV Aerial Photography Scenarios}},
  year = {2023},
  doi = {10.3390/s23167190},
}
Figure