Activity I.1 aimed to study existing in-house localization methods based on WiFi fingerprinting. Both acquisition methods and developed software platforms have been studied based on these methods. The proposed methods for processing the acquired data have also been studied. A single approach has been identified that proposes to automate the acquisition process by using a robotic platform. This is a less performing robotic platform that does not have Simultanous Localization and Mapping (SLAM) capabilities. Therefore, this aspect of the LOMAR project is an innovative approach in the field.

Activity I.2 aimed at studying existing neural network architectures that are suitable for location-based fingerprinting in the presence of noise. Studies have found that neural networks are still little used in this area. The identified work is briefly presented.

Activity I.3 aimed at analyzing and using the test data set available at CITST in preliminary tests with the available versions of neural networks. A neural network with 149/172 inputs, 4 hidden layers with 1200, 400, 100 and 20 neurons and ELU activation functions as well as a 2-out layer with linear activation functions was created.

Activity I.4 aimed at selecting the test location, defining test parameters, and manually purchasing data. The plans of several locations within the UPB were obtained. The final decision was to test along a corridor within the Faculty of Electronics because, being close to UPB’s premises, it allowed for repeated tests. During the stage, both software and hardware needed for data acquisition were developed.

Activity I.5 aimed at choosing the learning algorithms and setting the parameters of the models on the data set collected in I.4 activity. The data set has 3120 measurements in which 172 Wifi signal transmitters were captured. The following supervised learning algorithms have been chosen:

Activity I.6 aimed at:

Activity II.1 aimed to implement, integrate, and develop the testing methods of the robot hardware and software environments. In terms of hardware testing, two SLAM algorithms (Simultanous Localization and Mapping) have been considered along with various sensors on the Turtlebot2 robotic platform. Both the superiority of the RTAB-Map vs. Hector algorithm and the need to use as many sensors as possible to map spaces with few distinctive features (e.g., corridor mapping) have been found. In order to test the software, RSS WiFi data was acquired, which was used, together with data from LOMAR project collaborators, to develop and test machine learning based positioning algorithms. The comparison of learning models obtained by the LOMAR team with those of the collaborators revealed an increased accuracy of the LOMAR methods compared to those of the collaborators. Another important conclusion is the superiority of supervised learning methods, especially for large sets of data, which can be obtained by automating the collection process.

Activity II.2 aimed to creating a realistic scenario based on the “person tracking” technique, validating the location, and implementing a simplified “tracking” behavior. In this activity, a scenario was proposed (article submitted to “IEEE Robotics & Automation Magazine”) in which several collaborative robots collect data needed for localization and then provide services based on collected data. The YOLO2 network for tracking people was used. This is a method that produces invariant results in illumination and positioning. The method is capable of detecting people in different positions (standing, sitting, etc.). It’s also capable of detecting people within 3 meters of the robot’s camera. A disadvantage of this method is that it requires considerable computational power so that, depending on the capabilities of the robotic platform, it is potentially necessary to process the information on an external server and return the results to the robotic platform.

In this activity, mapping and navigation elements were implemented from the scenario presented in Activity II.2. Mapping the indoor environment using a robotic platform was done with both the Turtlebot2 robot and Tiago. Mapping with the Turtlebot2 robot was tested in the Leu building of the Faculty of Electronics and the Tiago in the Precis building of the Faculty of Automation. Various algorithms implemented in ROS packages have been tested. Also, a random walk algorithm has been implemented and tested for accuracy and robustness. The use of the Tiago robot has also been tested in the Gazebo simulation environment where crowded spaces were created to be mapped and then navigated.

Also, the purpose of this activity was to identify the intellectual property rights, which has been done by searching in EPO databse patents with the following themes:

1. internal positioning algorithms based on the received signal strength of the WiFi signal,
2. robots that work collaboratively to acquire data from WiFi signals,
3. an in-house learning machine based on the received WiFi signal strength signal,
4. neural networks for indoor positioning based on the signal strength of the received WiFi signal.

Activity II.4 aimed to:

1. publicly demonstrate the results, where members of both organizations UPB and CITST attended,
2. the dissemination of the project both through the creation and updating of the web page (http://www.lomar.pub.ro),
3. and through scientific publications: A publication was presented at the 12th Romanian International Conference on Communications, COMM 2018, and an ISI-indexed review article with Impact Factor 3.276, was submitted at IEEE Robotics & Automation Magazine.