Abstract
Catching waste in inland waters before it reaches the ocean is critical for environmental preservation. However, autonomous detection of floating debris presents unique challenges, including small target sizes, reflections, and environmental noise. This paper introduces FloW, the first comprehensive dataset and benchmark for floating waste detection from the perspective of Unmanned Surface Vehicles (USVs). FloW provides high-resolution imagery and synchronized millimeter-wave radar data to facilitate research in sensor fusion for robust aquatic object detection.
Introduction
Marine debris primarily originates from land-based sources such as rivers and canals. Once trash enters the open ocean, recovery becomes significantly more difficult as most of it sinks or enters the water column. While Unmanned Surface Vehicles (USVs) offer a scalable solution for intercepting waste in inland waters, they struggle with standard computer vision tasks due to the specific characteristics of aquatic environments.
Key challenges identified in this work include:
- Small Targets: Objects like bottles and cans occupy very few pixels (often less than 32x32).
- Reflections: Sunlight and bank-side structures create "ghost" objects on the water surface.
- Environmental Noise: Waves and floating vegetation can be easily misidentified as waste by standard models.
Methods and Datasets
The researchers developed a custom USV platform equipped with dual cameras—including a high-dynamic range (HDR) sensor—and a 77 GHz Frequency Modulated Continuous Wave (FMCW) radar.
The resulting FloW dataset is divided into two distinct components:
- FloW-Img: 2,000 images containing 5,271 labeled instances of floating waste. Over half of these are classified as small objects.
- FloW-RI (Radar-Image): 4,000 frames of synchronized camera and radar data. Radar provides robust spatial information (Range-Doppler Matrices and 4D Point Clouds) that remains effective despite visual glare or shadows.
Results and Discussion
Benchmarking various state-of-the-art models revealed that standard vision-only architectures struggle with the low signal-to-noise ratio of small debris on water. On the FloW-Img dataset, Cascade R-CNN achieved the highest accuracy with a mean Average Precision (mmAP) of 0.434, though it operated at a low frame rate of 3.9 FPS. In contrast, DSSD provided a more real-time capable performance at 28.6 FPS with an mmAP of 0.275.
The most significant improvement came from sensor fusion. In the FloW-RI multimodal tests, vision-only models achieved an $AP^{50}$ of 0.640. However, when fused with radar data using CRF-Net, the accuracy increased to 0.754. While radar alone is insufficient for identifying small debris due to lower resolution ($AP^{50}$ of 0.405), it serves as a critical "truth" source to filter out visual reflections and provide precise distance measurements.
Conclusion
FloW establishes a new frontier for aquatic environmental monitoring. By highlighting the limitations of current vision models in detecting small, reflective targets, the authors demonstrate that multimodal sensor fusion is essential for reliable autonomous river cleaning. Future work will focus on increasing category diversity and improving the real-time performance of fusion models to enable seamless navigation in complex, turbid water environments.
Summary updated