| Popis: |
One of the main problems in mobile robotics is to estimate the global position in complex symmetrical environments. Even when there are many devices or algorithms to achieve that goal, not all of them are useful in all kind of environments. GPS is typically used outdoors whereas algorithms based on Monte Carlo localization (AMCL) are used in- doors. However, they present some disadvantages. Thus, the GPS commercial devices do not work inside the buildings and the AMCL algorithms are limited in symmetrical envi- ronments for the fact that they needs to detect remarkable differences in the environment. Due to the mentioned limitations we propose a global localization approach for symmet- rical indoor environments based on the structure of topological maps. Here, geometrical and semantic information of static objects are considered, respectively, from LIDAR and RGB-D camera. Both sensors provide us, respectively, the information about occupancy areas and the scene perception. The proposed system is divided into four tasks. The first one is the classification of nodes according to their geometrical nature based on the LIDAR signature. The second stage is focused on object detection through a pretrained CNN based on YOLO (You Only Look Once) as model of convolutional neural network, which is able to work in real time. The third task corresponds to tracking and pose es- timation of objects, where is necessary the information from YOLO and the depth data from camera. Finally, the last task consists of estimating the robot’s global pose on the map from the output of object detector, their relative distance and their estimation pose. This algorithm compares the structure of detected nodes and objects with the structure defined on a reference annotated map. In order to match the degree of similarity of both structures we define a evaluation function and the highest value estimates the edge where is located the robot in the topological map. Our main contributions with respect to our previous work are the addition of depth of detected objects and the improvement of the evaluation function. |
