The 3D-CAVE project aims at the investigation of new embedded reconfigurable accelerators for the 3D mapping of underwater caves with new Visual-Inertial Simultaneous Localization and Mapping (VI-SLAM) algorithmic innovations to enable their use in extreme environments.
Project Overview
Groundwater, particularly that accumulated in karst aquifers, is one of the main freshwater resources that is used for agriculture, breeding, viticulture, arboriculture, industry, and human consumption. In Portugal, Maciço Calcário Estremenho (MCE) is one of the most important limestone massifs, not only due to its extension but also due to its diversified set of karst forms and springs that are used as the main water supply for several municipalities. Karst aquifers are a sensible and vulnerable ecosystem that is not only susceptible to pollution from industrial and domestic sources, it is also often affected by over-use endangering the surrounding environment.
One of the main tools that is used for the conservation and protection of these hydrogeological systems is the mapping and volumetric modeling of underwater caves created by the flow of water from karst aquifers. Such models provide researchers with precise knowledge of the underneath conduit network, in geographic position, depth, geomorphology, and seasonal/episodic water flow dynamics. Due to the underwater environment, the exploration, survey, mapping/characterization of flooded caves is traditionally done by specialized cave divers by resorting to manual depth, distance, and azimuth measurements.
Remote sensing (e.g., LiDAR), photogrammetry, and/or other machine vision methods, previously only used on land and aerial environments, are slowly making their way to underwater domains. In particular, Simultaneous Localization and Mapping (SLAM) and Visual Odometry (VO) techniques are used to track the pose of an agent in relation to its surrounding environment according to data from visual sensors. To accurately estimate the scale and structure of 3D objects and scenes, recent visual-inertial SLAM (VI-SLAM) approaches combine data from cameras and inertial measurement unit (IMU) sensors to enrich the robustness to drift and loss of localization problems and enable the generation of accurate volumetric models of the environment. However, adopting such techniques in flooded caves presents significant challenges, since they are hardly suited to operate over image sequences obtained in the conditions typically found in such environments, such as light refraction and absorption caused by water, lack of natural light, and presence of various forms of suspension underwater, which all may cause algorithms to lose tracking and localization. As such, innovative algorithmic improvements are required to enable the use of these techniques in underwater cave environments.
To enable their use, it is necessary to transport sensors through the underwater cave environment. Alternatively to the use of remotely-operated (or other) vehicles (ROVs), which may get stuck in complex rock formations and damage the cave, the required hardware can be easily attached to Diver Propulsion Vehicles (DPVs). These are commonly used by divers to move long distances underwater, and be used to passively collect data throughout the dive that can later be used for 3D reconstruction of the environment. However, since data is often only analyzed and processed with performance-demanding modeling tools in server machines, it may require multiple repetitive dives to fully map a single cave section. A possible solution is the deployment of specialized low-power VI-SLAM hardware accelerators based on embedded Field-Programmable Gate Array (FPGA) devices, with the goal of enabling on-location real-time model accuracy estimations to accelerate the process of surveying flooded karst systems.
Accordingly, this project aims at taking the first steps to develop a robust scientific platform for the volumetric characterization of underwater cave environments, to aid in the exploration, survey, and conservation of groundwater resources. To achieve such a goal, this project aims at developing a proof-of-concept laboratory accelerator prototype for real-time generation of 3D models from underwater cave footage and sensor data. To do so, it relies on a multi-disciplinary collaboration between academic partners and a national Non-Governmental Organization (NGO), that will undertake this challenge in three distinct fronts: 1) investigation of new techniques to enable the use of new VI-SLAM algorithms in underwater cave environments; 2) design and implementation of an FPGA-based accelerator prototype for real-time 3D reconstruction based on multi-camera/IMU sensor setups; 3) creation of a repository of underwater cave image sequences and sensor data obtained by specialized cave divers in the flooded karst networks of the Maciço Calcário Estremenho, to be used both as input data for algorithmic and prototype development. The outcomes of this exploratory project will be used as the base for the proposal of a subsequent project that will undertake the more ambitious challenge of developing a submersible device and fully functional underwater cave characterization framework.
Project Publications
