| T3.3.1 Intelligent AUV mapping |
| T3.3.2 AUV-ASV Cooperation |
Lead: UdG, Coronis, IQUA
Nowadays, AUVs carry out exploration missions by following pre-programmed lawn-mower trajectories while tracking the terrain to maintain optimal navigation and optimizing survey coverage and data quality. However, very little on-board processing of the collected data is performed, thus the AUV is not aware of its own performance in terms of the quality of the data that is being acquired, and mapping is often performed offline. State-of-the-art 3D mapping relies on sparse techniques such as bundle adjustment (BA) and structure from motion (SfM), which jointly estimate sparse 3D points of the scene and the vehicle poses. To obtain an optimal solution, a nonlinear optimization is performed on a complete set of vehicle poses and 3D points observed by the camera in a stochastic estimation framework which accounts for Gaussian noise models in the observations. However, for large scale applications, this is an expensive procedure and it is normally performed offline, after the acquisition process. This post-mission processing leads to gaps in the coverage or detection of interest areas that have not been properly covered, and despite the best efforts, several deployments over the same area are still commonly required, significantly rising the total cost of the mission due to the costs of ship-time and highly trained personnel.
Within Task T3.3.1, algorithms will be developed to endow the explorer AUV with more intelligence by means of the following strategies: (1) the robot will process the collected data onboard to estimate the quality of the acquired data in mission time; (2) a machine learning process running onboard will recognise interest targets, e.g. seagrass meadows, amphorae field, shipwreck artefacts, etc. in mission time; (3) once an interest target is identified, the AUV will re-plan its mission to cover the target area with higher resolution at a shorter distance and approach the target from optimal viewpoints; (4) once an interest area is defined, the AUV will build a sparse optical map using a stochastic estimation framework, while the images are being acquired. The AUV will monitor which parts of the environment have been adequately mapped according to relevant criteria – namely the 3D model complexity and geometric uncertainty-, addressing at mission-time the question “which parts of the environment shall we map next or re-visit?”
The project aims to increase the TRL of such a system from small scale prototype (TRL4) to demonstration system (TRL 6-7). This will be achieved by solving the BA problem in an efficient way, during the mission, in an incremental fashion. We have derived a new formulation10* that is able to provide not only the solution of the optimization (location of the robot and 3D points in the environment), but also the value of the associated uncertainty of the 3D reconstruction at mission-time. The ability to obtain the reconstruction and associated quality estimation during the time of the mission subsequently enables the possibility of active feedback about the quality of the acquired data (and the 3D structure) as well as the identification of poorly mapped or even missing areas. This crucial information will have applications not only for autonomous vehicles, but also for teleoperated ones. Endowed with this additional knowledge, skilled pilots and autonomous planning schemes will be able to alter the mission in progress. By guiding the vehicle towards the problematic areas, the quality of the final representation will be significantly improved. Furthermore, such quality- aware surveys will not only reduce the need for additional deployments of the vehicle but by utilizing the next best view suggestions also greatly improve the survey efficiency, further reducing the mission time and costs.
