YOLO-Drone:Airborne real-time detection of dense small objects from high-altitude perspective

Zhu Li, Jiahui Xiong, Feng Xiong, Hanzheng Hu, Zhengnan Jiang

TL;DR

This paper introduces YOLO-Drone, a novel real-time object detection algorithm specifically designed for Unmanned Aerial Vehicles (UAVs) to tackle the challenge of detecting small, dense objects from high altitudes. It achieves state-of-the-art performance on benchmark datasets and demonstrates exceptional capabilities in challenging low-light conditions, particularly when using a specialized silicon-based golden LED light source.

What Problem Does This Paper Address?

Detecting small objects in aerial imagery captured by drones is a significant challenge in computer vision. These objects often suffer from low resolution, image degradation, and can be densely packed or obscured. Existing object detection methods, while advanced, struggle with the unique characteristics of UAV imagery, leading to missed detections and inaccurate localization. This is particularly problematic for critical applications like surveillance, traffic monitoring, and disaster management, where reliable detection of small targets is paramount. The paper also highlights the difficulties of night-time detection due to poor illumination and light pollution.

Methods

The core of this work is the YOLO-Drone algorithm, which builds upon the YOLO architecture. Key innovations include:

  • Darknet59 Backbone: A new backbone network is introduced, deepening the architecture and incorporating a four-scale detector. This design aims to better capture high-frequency information from small objects.
  • MSPP-FPN Neck: A novel feature aggregation module, MSPP-FPN (Multi-Scale Spatial Pyramid Pooling - Feature Pyramid Network), is proposed. It combines a Spatial Pyramid Pooling (SPP) module with three Atrous Spatial Pyramid Pooling (ASPP) modules. This enhances the network's ability to aggregate multi-scale features and improve receptive field coverage.
  • GIoU Loss Function: The Generalized Intersection over Union (GIoU) loss is adopted for bounding box regression. This is a departure from traditional Mean Squared Error (MSE) loss and is shown to improve localization accuracy by considering the overlap between predicted and ground truth boxes, even when they don't intersect.
  • K-means Clustering for Anchors: The paper utilizes a k-means clustering approach with Intersection over Union (IoU) scores to determine optimal anchor box sizes tailored to the specific datasets, further improving detection for objects of varying scales.

Datasets

The YOLO-Drone algorithm was evaluated on three datasets:

  • UAVDT: A benchmark dataset for UAV-based vehicle detection and tracking, featuring complex scenarios.
  • VisDrone2021-DET: A large-scale aerial image dataset for object detection, including various object categories and challenging conditions.
  • UAV-LED-G: A newly introduced private dataset collected at night under a silicon-based golden LED light source. This dataset is specifically designed to evaluate performance in low-light conditions and features small objects like pedestrians.

Key Insights and Ideas

  • The Importance of Specialized Lighting: The research strongly suggests that for night-time object detection from UAVs, the light source itself is a critical factor. The silicon-based golden LED offers significant advantages over traditional light sources by reducing light pollution and improving penetration in adverse weather, leading to much higher detection accuracy.
  • Multi-Scale Feature Fusion is Crucial: The MSPP-FPN module's success highlights the necessity of effectively combining features from different scales to capture both global context and fine-grained details of small objects.
  • GIoU for Better Localization: Moving beyond simple IoU or MSE, the adoption of GIoU loss demonstrably improves the precision of bounding box regression, which is vital for accurately identifying small targets.
  • Tailoring Anchors Matters: Using k-means clustering to adapt anchor box sizes to specific datasets significantly improves the model's ability to detect objects across different scales.

Key Results

  • State-of-the-Art Performance: YOLO-Drone outperforms existing SOTA methods on the UAVDT dataset, achieving a maximum mAP of 34.04% with a real-time inference speed of 53 FPS. It also shows significant improvements on the VisDrone dataset.
  • Exceptional Night Detection: On the custom UAV-LED-G dataset, YOLO-Drone achieves an impressive mAP of 87.71%, significantly surpassing YOLO series performance under ordinary light sources.
  • Robustness in Complex Scenarios: The algorithm demonstrates superior performance across various challenging scenarios, including different weather conditions, viewing angles, and altitudes.
  • Efficiency: The 53 FPS inference speed on the NVIDIA RTX 2080ti GPU meets real-time requirements for UAV applications.

Limitations and Future Work

  • Low-Altitude Performance Degradation: The paper notes a slight decrease in mAP at very low altitudes, potentially due to class imbalance in the dataset and the mismatch between large objects and anchor frames.
  • Dataset Specificity: While the UAV-LED-G dataset is a valuable contribution, further expansion and diversity in night-time datasets would be beneficial.
  • Future Work: The authors suggest that YOLO-Drone is a highly effective solution for UAV object detection, particularly for night tasks. The developed system has been integrated into autonomous cruise UAV platforms, indicating potential for real-world deployment in security, transportation, and intelligent city applications. Further research could explore more advanced feature fusion techniques or adapt the model for even more challenging environmental conditions.

Summary updated

Abstract

Unmanned Aerial Vehicles (UAVs), specifically drones equipped with remote sensing object detection technology, have rapidly gained a broad spectrum of applications and emerged as one of the primary research focuses in the field of computer vision. Although UAV remote sensing systems have the ability to detect various objects, small-scale objects can be challenging to detect reliably due to factors such as object size, image degradation, and real-time limitations. To tackle these issues, a real-time object detection algorithm (YOLO-Drone) is proposed and applied to two new UAV platforms as well as a specific light source (silicon-based golden LED). YOLO-Drone presents several novelties: 1) including a new backbone Darknet59; 2) a new complex feature aggregation module MSPP-FPN that incorporated one spatial pyramid pooling and three atrous spatial pyramid pooling modules; 3) and the use of Generalized Intersection over Union (GIoU) as the loss function. To evaluate performance, two benchmark datasets, UAVDT and VisDrone, along with one homemade dataset acquired at night under silicon-based golden LEDs, are utilized. The experimental results show that, in both UAVDT and VisDrone, the proposed YOLO-Drone outperforms state-of-the-art (SOTA) object detection methods by improving the mAP of 10.13% and 8.59%, respectively. With regards to UAVDT, the YOLO-Drone exhibits both high real-time inference speed of 53 FPS and a maximum mAP of 34.04%. Notably, YOLO-Drone achieves high performance under the silicon-based golden LEDs, with a mAP of up to 87.71%, surpassing the performance of YOLO series under ordinary light sources. To conclude, the proposed YOLO-Drone is a highly effective solution for object detection in UAV applications, particularly for night detection tasks where silicon-based golden light LED technology exhibits significant superiority.
BibTeX
@misc{li2023yolo,
  author = {Zhu Li and Jiahui Xiong and Feng Xiong and Hanzheng Hu and Zhengnan Jiang},
  title = {{YOLO-Drone:Airborne real-time detection of dense small objects from high-altitude perspective}},
  year = {2023},
  eprint = {2304.06925},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.06925},
}
Figure