[Review] 3D Cooperative Coverage

March 27, 2019 ยท 4 minute read

Article Review 27-03-2019

Cooperative Coverage for Surveillance of 3D Structures

Adaldo, Antonio, et al. “Cooperative coverage for surveillance of 3D structures.” 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2017.

Aims of the paper

The authors aim to propose a new planning algorithm for coverage of complex 3D structures with a network of multiple anisotropic sensing agents. They look specifically at what they term the static sensor placement problem where they look for optimal points for which information gained is maximum, rather than the dynamic sensor placement problem for coverage while in motion.

Paper Summary

Many previous works consider portions of the aim of this work. For instance there has been a lot of work for omnidirectional sensors where Voronoi diagrams are used. In cases where work uses anisotropic (non-omnidirectional) sensing, there is not usually an attempt to understand the morphology of the structure. Also there are very works which focus on the coverage in 3D space.

Their method uses the notion of voronoi tessellation - their 3D generalisation of 2D voronoi covering - to consider the problem of static coverage of 3D structures. They require this notion as for anisotropic sensing, the sensing pattern not only depends on sensor to surface alignment, but also between the sensor and surface normal. Their method is also capable of coverage on the fly with communication between the different agents. A notable point is that the sensors generate a set of landmarks as opposed to a full point cloud of a structure to discretise the continuous environment. A collision free, optimal placement path is found by minimising a cost function which involves the voronoi tessellation, in effect maximising the number of landmarks an agent can see while reducing collision.

The begin by mathematically laying out the problem, explaining their coverage score function as the number of landmarks visible and describing voronoi tessellation as a tuple of landmarks, sensors, and landmark-sensor mappings for which each sensor is observing the maximum possible landmarks it can. The goal is to optimise the coverage function such that the system of agents reaches a voronoi tessellation. Then the control algorithm follows the following steps:

  1. Each sensor is abstracted to a pose and sensor footprint
  2. Control law is based on a gradient ascent motion on its own coverage w.r.t change in landmark assignments
  3. Collision Avoidance is included as a safety radius in coverage function
  4. The landmark’s found are down-sampled to a desired level-of-detail for ease of communication
  5. Each agent uses the KD-Tree algorithm combined with PCA to estimate the surface norms described by its set of landmarks to produce a full map.

The system is tested in software simulation using multiple ROS nodes to simulate a distributed set of agents, each with monocular camera sensing footprints. This showed incredibly promising results with an even distribution of landmarks from the structure across the agents. A sensible communications procedure allowed almost complete coverage to be achieved by all agents. The authors also tested this system in reality using 2 AscTec Neo Hexacopters with cameras attached. Pose-estimation was achieved via Motion Capture equipment, with an EKF and MPC used as the on-board controller. This produced alright results with coverage being shown to be attained, however trajectories turned out to be erratic and it is not clear if coverage was evenly split between the agents.

Paper Review

I felt that this was a good paper that contributed some novel work to the field. As mentioned before, little work has been done for 3D cooperative coverage and this work does well to be theoretically thorough while providing some results in real experiments and trials. the generalised notion of Voronoi Tessellation, although conceptually simple, could be applied in many situations in different areas of research, including my own. The author’s appear to be fairly thorough with their consideration of issues, going to include both obstacle avoidance and inter-agent communication for robustness and convergence of maps. I was particularly impressed with their mathematical formulation which gave me a lot of insight into the problem.

It was a shame to see the lack of solid results in the real-life test, after the seeming success of the software simulation. The paper glosses over the results, but they do seem a bit disappointing, I hope to see some further results, possibly if they have more time to test. I also wished they went into more detail about how they conducted inter-agent communication as little detail is given. It would have been interesting to see what would have happened if communication was lost or a drone was disabled, would the agents self-organise and continue anyway.

Overall a fairly good paper. Its well written and well organised with a decent literature review and clearly explained methodology and experimentation. The use of mathematics is contained and does not over complicate the explanations at all and the graphs and diagrams are informative and not superfluous. I would recommend this paper to people who are researching either building inspection methods, or multi-agent coverage methods.