Unlike images, which are structured as regular 2D or 3D grids, point clouds are unstructured 3D data consisting of discrete coordinates acquired using depth sensors or computer vision techniques. While their processing involves challenges such as acquisition noise, occlusions, irregular sampling, and large data volumes (sometimes reaching billions of points), point clouds are now widely used in various domains, including robotics, virtual or augmented reality, cultural heritage, and urban planning.

Our contributions in point cloud processing

The Image Processing and Pattern Recognition group has been conducting research in point cloud processing since 2022, focusing on the following tasks:

• Methodology: The team works on adapting deep learning architectures to address the unstructured nature of point clouds, with an emphasis on the computational efficiency of the developed methods. The main focus is on improving semantic segmentation and object detection in 3D scans.
• Applications: The primary application of the team research in point cloud processing is underwater 3D scene reconstruction using exploration robotics. This involves drones designed for both autonomous and remotely operated navigation in marine and underwater environments. These small, easily deployable drones are capable of data acquisition in challenging underwater conditions. The team focuses on enhancing data acquisition synchronization from multiple viewpoints and improving contextualization using sensors for attitude, depth, and timestamp measurements [2]. After data acquisition, the team develops high-precision 3D reconstructions and automatic species identification [1], supporting valuable applications in marine exploration, mapping, and environmental monitoring.

Related Publications