The task of UAV-based maritime rescue object detection faces two significant challenges: accuracy and real-time performance. The YOLO series models, known for their streamlined and fast performance, offer promising solutions for this task. However, existing YOLO-based UAV maritime rescue object detection methods tend to prioritize high accuracy, often at the expense of real-time performance and ease of implementation and expansion. This study proposes a modular plug-and-play optimization approach based on the YOLOv8 framework, aiming to enhance real-time performance while maintaining high accuracy for UAV maritime rescue object detection. The proposed optimization modules are flexible, easy to implement, and extendable. In experiments on the large-scale publicly available SeaDronesSee dataset, our method achieved a 13.53% improvement in accuracy over YOLOv8x while reducing computational cost by 85.63%. Additionally, it surpassed the detection speed of the SeaDronesSee official code’s two-stage detector by over 20 times, while maintaining comparable accuracy. Furthermore, our analysis of the experimental results highlights differences in detection difficulty among various objects and potential biases within the dataset.
Accurately and rapidly detecting objects and their locations in drone-captured images from maritime search and rescue scenarios provides valuable information for rescue operations. The YOLO series, known for its balance between lightweight architecture and high accuracy, has become a popular method among researchers in this field. Recent advancements in the newly released YOLO11 model have demonstrated significant progress in general object detection tasks across everyday scenarios. However, its application to the specific task of drone-based maritime search and rescue still leaves substantial room for improvement. To address this gap, we propose targeted optimizations to enhance YOLO11’s performance in this domain. These include integrating a Space-to-Depth module into the Backbone, incorporating a content-aware upsampling algorithm in the Neck, and adding an extra detection head to better exploit shallow image features. These modifications significantly improve the model’s ability to detect small, overlapping, and rarely occurring objects, which are common challenges in maritime search and rescue tasks. Experimental evaluations conducted on the large-scale SeaDronesSee dataset demonstrate that the proposed optimized YOLO11 outperforms YOLOv8, YOLO11, and MambaYOLO across all scales. Moreover, under lightweight configurations, the model achieves substantial performance gains over YoloOW, a method renowned for its accuracy but depends on heavyweight configurations. In the lightweight complexity range, the proposed model achieves a relative accuracy improvement of 20.85% to 43.70% compared to these state-of-the-art methods. The code supporting this research is available at https://github.com/bgno1/sds_yolo11.
The specificity of scenarios and tasks in Unmanned Aerial Vehicles (UAV)-based maritime rescue poses challenges for detecting targets within images captured by drones in such environments. This study focuses on leveraging heuristic methods to extract data features from specific UAV maritime rescue images to optimize the generation of anchor boxes in detection models. Experiments conducted on the large-scale SeaDronesSee maritime rescue dataset, using the MMDetection object detection framework, demonstrated that the optimized anchor boxes, improved model performance by 48.9% to 62.8% compared to the framework's default configuration, with the most proficient model surpassing the official highest SeaDronesSee baseline by over 49.3%. Further analysis of the results revealed the variation in detection difficulty for different objects within the dataset and identified the reasons behind these differences. The methodology and analysis presented in this study hold promise for optimizing UAV-based maritime rescue object detection models as well as refining data analysis and enhancement.
Accurate object detection in drone-captured images is crucial for maritime search and rescue missions. To address the challenges posed by sample constraints and data imbalance in existing datasets, we introduce a greedy algorithm-based dataset rebalancing method tailored for the SeaDronesSee v1 and v2 datasets. Our method aims to alleviate potential sample constraints by redistributing samples between the training and validation sets. To validate the effectiveness of our approach, we trained and evaluated four state-of-the-art object detection models—Faster Region-based Convolutional Neural Network (Faster R-CNN), Cascade R-CNN, You Only Look Once version 8 (YOLOv8), and YOLO11—on both the original and rebalanced versions of the datasets. The results demonstrate that dataset rebalancing significantly improves the performance of all models, with YOLO models benefiting more than their two-stage counterparts. Our contributions include analyzing the sample constraints in the SeaDronesSee series datasets, proposing a greedy algorithm-based rebalancing strategy, and validating its effectiveness through experiments. This study optimizes the sample distribution within each dataset, enhancing model training and generalization, and provides valuable insights for addressing data scarcity issues in similar scenarios. The code supporting this study is available at https://github.com/bgno1/sds_rebalance.