RESUMEN
Robotic missions for solar farm inspection demand agile and precise object detection strategies. This paper introduces an innovative keypoint-based object detection framework specifically designed for real-time solar farm inspections with UAVs. Moving away from conventional bounding box or segmentation methods, our technique focuses on detecting the vertices of solar panels, which provides a richer granularity than traditional approaches. Drawing inspiration from CenterNet, our architecture is optimized for embedded platforms like the NVIDIA AGX Jetson Orin, achieving close to 60 FPS at a resolution of 1024 ×1376 pixels, thus outperforming the camera's operational frequency. Such a real-time capability is essential for efficient robotic operations in time-critical industrial asset inspection environments. The design of our model emphasizes reduced computational demand, positioning it as a practical solution for real-world deployment. Additionally, the integration of active learning strategies promises a considerable reduction in annotation efforts and strengthens the model's operational feasibility. In summary, our research emphasizes the advantages of keypoint-based object detection, offering a practical and effective approach for real-time solar farm inspections with UAVs.
RESUMEN
Counting objects in images is a very time-consuming task for humans that yields to errors caused by repetitiveness and boredom. In this paper, we present a novel object counting method that, unlike most of the recent works that focus on the regression of a density map, performs the counting procedure by localizing each single object. This key difference allows us to provide not only an accurate count but the position of every counted object, information that can be critical in some areas such as precision agriculture. The method is designed in two steps: first, a CNN is in charge of mapping arbitrary objects to blob-like structures. Then, using a Laplacian of Gaussian (LoG) filter, we are able to gather the position of all detected objects. We also propose a semi-adversarial training procedure that, combined with the former design, improves the result by a large margin. After evaluating the method on two public benchmarks of isometric objects, we stay on par with the state of the art while being able to provide extra position information.