• a system based on industrial cameras, visual tags mounted on the robot, and ad hoc software
• a system using a set of Laser Range Finders andad hoc software, used to localize a rectangular hull mounted on the robot

Both systems were set up to cover a subset of the explored environment (we called that region of space the “GT area”). The two trajectory-reconstruction systems worked independently one from the other (and from the sensor systems on board of the robot). Off-line, their recorded outputs were combined to generate a third stream of ground truth data with still lower error. Such combined data stream, describing the trajectory of the robot whenever it traversed the GT area, is the final one, supplied as ground truth with Rawseeds’ datasets.

The vision-based ground truth collection system was based on a multiple-camera GigE Vision system composed of Prosilica GC-series cameras, a Linux-based PC running custom software and a Gigabit Ethernet switch to concentrate all the vision data on a single network link to the PC. The LRF-based system used well-established Sick LMS200 sensors, and again a Linux-based PC running custom software.

A quantitative analysis of the performance of Rawseeds’ indoor trajectory reconstruction setup can be found in Deliverable D2.1 – part 2.

Base station for the RTK-GPS system

• a Trimble 5700 GPS receiver with Zephyr GPS antenna, mounted on the mobile robot (rover);
• a Trimble MS750 receiver with Zephyr antenna, mounted in a fixed location (base station);
• a Trimble PDL450 radiomodem, to transmit positioning information from the base station to the rover.

This system has – in best conditions of satellite signal reception – a positioning precision of 1cm horizontal and 2cm vertical. In all the locations where data acquisition occurred, the base station has been set in elevated points to maximize satellite visibility, i.e. on the top of tall buildings with no obstructions in their vicinity.

In actual use, satellite visibility has never been a problem for the base station, while it proved to be extremely critical for the rover. The presence of buildings, structures or even trees in the vicinity of the rover heavily disturbed the reception of satellite signals. Despite careful setup and planning of the acquisition sessions with reference to the visibility of satellite constellations (which evolve quickly through the day, and change slowly from day to day), RTK-GPS technology in urban settings proved to suitable for Rawseeds’ requirements, but very critical and sensible to setup issues. Even in quite sparse urban environments, such as the Polimi-Durando location, the best positioning quality (corresponding to the so-called RTK fixed integers mode of operation) was available only on limited parts of the robot’s trajectory. This was expected, but we hoped for a wider tolerance. Lower-grade positioning precisions (within a few centimeters of error, so still compliant with Rawseeds’ specifications) were available through significant parts of the robot’s path: therefore it has been possible for Rawseeds to gather good ground truth data in outdoor environments. However, our experience showed that RTK-GPS technology is too costly and prone to setup and reception issues to be considered as a not-auxiliary sensor in robotics applications.

A quantitative analysis of the ground truth data provided by the RTK-GPS system of Rawseeds is available in Deliverable AD2.3.

• the trajectory of the robot during (parts of) the data-gathering session
• the executive drawings describing the environments explored by the robot

In addition to the above, each dataset includes a list of positions of suitable features extracted from the executive drawings. These position are used when applying some of the performance metrics specified by the Benchmark Problems.
Ground truth data have been collected by means that are separate from the sensors mounted aboard the robot, in order to be completely independent from the datasets included in the BPs. In the case of trajectories, ground truth information are time-sinchronized with the datasets.