Research


Summary

The project aims at improving several key matters regarding Security based on Multi-Robot Systems.

Thus, several work-packages were defined in order to organize the work to carry out.

Requirement analysis for Multi Robot Systems devoted to perform security missions.

As a result of this work, a document was created. Main requirements were settle and analyzed for different security applications. (only available in Spanish)

Test Environment Modeling and Simulation

In order to allow fast prototyping, Webots  artificial environments have been designed. Robots modules have been integrated by using ROS (Robot Operative Systems) which allows integrating virtual environments and real robots and sensors.

   

 

UAV's and UGV's Control and Integration

Several commercial robots have been integrated in a Multi-Robot fleet by using ROS, which allows sharing information, fast prototyping and provides control architecture for the whole system.

The robot integrated so far are:

Cooperation among Aerial and Ground Vehicles

On this regard, three sub-areas have been settled down:

- Detection

Detection of mobile objects is a key issue in the development of the project. Algorithms and techniques for detection from both aerial and ground vehicles have been developed. Multi-robot detection algorithms have been created in order to share a unique list of dynamic object among all the robots in the fleet.

Detection algorithms are in charge of analyzing the information from sensors (cameras and lasers) in order to estimate the position, size and color of the dynamic object in the surroundings. Three kind of algorithms have been developed; two of them for ground vehicles and another for aerial ones.

Unlike detection algorithms for ground vehicles that use the laser readings in order to generate a 3D model from the environment, algorithms for aerial vehicles use a video camera (there are strong restrictions on its size and weight). Frames from this camera are processed in order to obtain optical flow so as to be compared with the estimated due to the drone movement.


 

 

As a part of Detection, a noteworthy efort is devoted to object identification. Thus, algorithms por humans and robots recognition are being developed and tested on real and virtual scenarios by using both Laser and Camera information.


 

 

 

- Tracking

There is no doubt that tracking algorithms are the basis for a good performance in security task. The objective of these algorithms is to perform matching between the information obtained from detection algorithms and the list of dynamic object that each robot  keeps. These algorithms are based on Mahalanobis and Bhattacharyya distances. A state vector into a Extended Kalman Filter is used in order to perform the tracking between detected and tracking objects.

Videos of tracking will be available shortly.

- Localization and Mapping

Apart from conventional GNC modules for UAV and UGV based on INS-GPS integration and odometry for ground vehicles, several proposals for cooperative navigations are being developed in the project. The following research lines are currently active:

- Multi-Robot Mapping.

Techniques of Global and Local maps stored and shared among the robotic team are applied in order to perform a global mapping.

- Cooperative localization based on visual recognition.

This technique aims at improving the position estimation of a robot when another member of the robotic team that rely on better position estimation is visually detected. So as to perform the visual recognition, a previous registration model should be carried out.

- Cooperative Air-Ground navigation and path planning.

The Navigation and obstacle avoidance is executed in different steps. Thus, the UGV identifies the obstacles that are in its path into the aerial map and then some techniques based on potential fields are applied in otder to perform a reactive navigation.

 

 

   

 

Cooperative multirobot surveillance.

It is focused on the problem of performing multirobot patrolling for infrastructure security applications in order to protect a known environment at critical facilities. Thus, given a set of robots and a set of points of interest, the patrolling task consists of constantly visiting these points at irregular time intervals for security purposes. Current existing solutions for these types of applications are predictable and inflexible. Moreover, most of the previous work has tackled the patrolling problem with centralized and deterministic solutions and only few efforts have been made to integrate dynamic methods. Therefore, one of the main contributions of this work is the development of new dynamic and decentralized collaborative approaches in order to solve the aforementioned problem by implementing learning models fromGame Theory. The model selected in this work that includes belief based and reinforcement models as special cases is called Experience. The problem has been defined using concepts of Graph Theory to represent the environment in order to work with such Game Theory techniques.

 

Multi  Aerial Vehicles Surveillance.

A fleet of aerial robots,  has to cover a interest area to be monitored by following  continuous and smooth trajectories while avoiding obstacles or prohibited areas and revisiting the minimum number of points. Furthermore, since not all areas are suitable for taking off or landing, the trajectory must ensure starting and ending points in places that fulfill all the requirements (safety margins, space enough for operation, capability for picking up and dropping off, accessibility).

In order to properly plan the task that an aerial robot has to perform in surveillance missions of large outdoor areas, the motion planning subfield known as Coverage Path Planning (CPP) is addressed.

Motivated by the demand for this type of planning in our past and present projects, a formal analysis of the CPP problem has been performed, in particular emphasizing a grid-based approach where a priori map of the environment is available. This map is approximately decomposed into cells, the grid-based approaches are feasible solutions when the environment is static and well known in advance, and also when precision and processing time are not in issue.The presented methodologies have not only been simulated but also applied in real systems, where their effectiveness has been validated.