Research
Research
Master Thesis
Supervisor
Full-Time Principal Researcher B in the Computer Science Department at the Instituto Nacional de Astrofisica Optica y Electronica (INAOE) based in Puebla City, in Mexico.
Honorary Senior Research Fellow in the Computer Science Department at the University of Bristol, in the UK.
Abstract
The Global Position System (GPS) has become an essential sensor for applications and public systems, maritime systems, and aerial vehicles in the robotics area. Traditionally, autonomous flight in outdoor areas is possible thanks to GPS devices that enable the Unmanned Aerial Vehicle (UAV) to obtain its position in latitude and longitude coordinates. However, GPS may become unreliable when the drone flies in environments where the signal may get occluded. Malicious attacks may also compromise the GPS signal, aiming to block the signal or replace it with spurious data. Motivated by these scenarios, the proposed approach relies on a methodology to estimate the GPS position of a UAV using Convolutional Neural Networks (CNN) and a learning-based strategy. For the latter, we adopted two learning scenarios: 1) offline learning; 2) online learning where we address the re-localisation problem and geo-localisation in a scenario where GPS devices fail.
We argue that our approach could be used as a backup system to return the UAV home when the GPS device is not working. We performed performance tests with aerial images and videos captured with the Drone Matrice 100 and Drone Parrot Bebop 2 in two scenarios with different trajectories to demonstrate our approach using a compact CNN and online training implementation. The experiments present suitable results, on average CompactPN obtained an error of 2.60 to 6.16 metres with a speed of 107.88 fps. Likewise, our implementation based on the AR1* method obtained an accuracy of 0.7931 to 0.9282 with a performance speed of 127.33 fps for online training results.
This research produced the following list of publications:
Index Journals:
Aldrich A. Cabrera-Ponce, J. Martinez-Carranza. “Convolutional Neural Networks for Geo-Localisation with a Single Aerial Image”. Journal of Real-Time Image Processing. February, 2022. JCR Q3. https://doi.org/10.1007/s11554-022-01207-1
Peer-Reviewed International Conferences:
Aldrich A. Cabrera-Ponce, and J Martinez-Carranza. “Aerial Geo-Localisation for MAVs using PoseNet”. IEEE 5th Workshop on Research, Education and Development of Unmanned Aerial Systems (RED-UAS). IEEE, Cranfield, UK. November 2019.
Aldrich A. Cabrera-Ponce, Manuel Martin-Ortiz, and Jose Martinez-Carranza. "Continual Learning for Multi-camera Relocalisation". 20th Mexican International Conference on Artificial Intelligence (MICAI). Mexico City, Mexico. October 2021.
Engineering Thesis
Supervisor
Full-Time Principal Researcher B in the Computer Science Department at the Instituto Nacional de Astrofisica Optica y Electronica (INAOE) based in Puebla City, in Mexico.
Honorary Senior Research Fellow in the Computer Science Department at the University of Bristol, in the UK.
Abstract
Recently, unmanned aerial vehicles (UAVs) have grown in the research areas of aerial robotics, thanks to their on-board cameras. The videos captured with the cameras allow the exploration of environments for solutions such as object detection, surveillance and those requiring aerial photography for disaster inspection. However, such a task requires a correct joining of the images, in addition to real-time. The present thesis aims to perform the stitching of aerial images to generate a real-time mosaic, based on visual descriptor correspondences extracted from the images. The first stage consisted of the calculation of the ORB descriptors and their comparison based on the libraries of a SLAM system. The second stage consisted of joining these images by calculating their transformation matrix and their correspondences. These correspondences are important to show those points that coincide between a first image and the second one, combining them in a wide image in such a way that a mosaic is generated with the joined images. The system was tested in four outdoor environments with image capture at different heights and the results show that the system achieves the correct joining of the images to generate a mosaic in real-time.