Page - 159 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

4.6. Locateuser This function requires navigation to a set of predefined searching positions which in this case were usually located in themiddleof the roomsorwere thesameas thecallbuttonplaces. Sincedetecting the user while the robot is rotating is hard, several shorter rotations were performed, stopping to call and detect the user in between. Using discrete motion rotations with our settings resulted in a more abrupt movement that affected localization, so a navigation task including only an orientation subtask was preferred. Still, the orientation estimate was sometimes not so good after the short rotations and the uncertainty associated to the orientation measurement had to be adjusted. Also, it was usually better to define this kind of positions in places where distinctive references were present and in relatively open space so that localization could get better before reaching narrower areas that requirebetter accuracy. 5. RGB-Dbasednavigation in homeenvironments One of the first things to check with the proposed sensor setup, depending on the environment, is that the height interval to be considered for the bottom sensor virtual scan generation may need to be changed. In general, we used a fixed width of 8 cm around the horizontal plane at the sensor’s height. In the presence of low sofas and similar furniture, however, this interval had to be lowered since otherwise the virtual scan generated with the largest measurements included irregular surfaces suchascushions, etc. Fig. 3 illustrates thiskindofproblem. Figure 3. If the height interval considered for the bottom virtual scan is too high in the presence of low sofas, irregularborders andcushionswillbepresent in the generatedscan. Regarding complete map building, the main limiting factors are the small field of view and the range properties of the RGB-D sensors. Since home environments may have narrow areas and excessive rotationsduring themappingprocesscan lead to lessaccurate resultsandevencausedistortion, some wallsmaybemissing in theresultingmaps. This fact, togetherwith thepossibility thatcurrent sensor data overwrite the global map, can bring about problems in global path planning (see Fig. 4 for an example). One first thing that can be done to prevent this problem when the robot is not facing the obstacle is carefully edit the map manually, to include the complete walls without risking the map building processoverall result. This,however,doesnothelpwhenthecurrentsensordataareusedtooverwrite andclear themap. Inthatcase,withthisfeatureincludedwithintheMIRAbasedimplementation,one option is to add MIRAnogoareas in order to avoid undesired plans when the robot is too close to a wall. Theproblemis that inverynarrowareassomemarginmustbeallowedforpossible localization 159