Por favor, use este identificador para citar o enlazar este ítem:
http://cimat.repositorioinstitucional.mx/jspui/handle/1008/1150| Crowd Analysis with Deep Learning | |
| Javier Gonzalez-Trejo | |
| Acceso Abierto | |
| Atribución-NoComercial | |
| Ingeniería del software | |
| This thesis tackles the problem of vision-based crowd analysis, with one or multiple monocular cameras and using modern deep learning techniques. To advance in the research of Crowd analysis applications, in this thesis, we develop solutions to real problems based on our own lightweight Deep Learning density map generator to solve recent challenges that involve the detection and counting of crowds in monocular cameras. For that, first of all, identify the state-of-the-art crowd counting and detection techniques, as well as the most important public datasets that serve as the basis for this work and future developments. More precisely, we will cover both, the development of a lightweight density map generator for real-time embedded applications, and the development of crowd analysis applications using state-of-the-art density map generators to solve two interesting real-world problems: 1) Safe Landing Zones detection in populated areas for Unmanned Aerial Vehicles (UAVs); and 2) Automatic Social Distance Monitoring. The lightweight density map architecture was developed from a pruned density map generator to reduce the number of parameters, and trained using the Bayesian Loss to improve its accuracy in the crowd counting and detection tasks, obtaining state-of-the-art Mean Square Error accuracy compared with the literature, while maintaining a competitive number of parameters. Using the lightweight density map generator, two crowd analysis solutions were developed. More specifically, we developed a Safe Landing Zones detection and tracking algorithm for UAVs emergency landing in populated scenarios, using the lightweight density map algorithm embedded in a device with limited computational resources, considering a camera mounted in the UAV. The density map is used to generate an occupancy-free mask projected to the so-called head plane, where the biggest circles free of people are found and tracked using Kalman filters and the Hungarian algorithm for data association. The safe landing detection algorithm was tested under real scenarios using data recorded from a drone, showing promising results. On the other hand, an Automatic Safe Distance Monitoring framework to train Deep Neural Networks using density maps from non-social distance conforming crowds was also developed in attention to the recent COVID-19 pandemic outbreak. Based on public available crowds datasets, we propose a density map and a segmentation-based solution, demonstrating superior perfo | |
| 20-08-2021 | |
| Trabajo de grado, maestría | |
| OTRAS | |
| Versión aceptada | |
| acceptedVersion - Versión aceptada | |
| Aparece en las colecciones: | Tesis del CIMAT |
Cargar archivos:
| Fichero | Descripción | Tamaño | Formato | |
|---|---|---|---|---|
| ZAC TE 77.pdf | 21.75 MB | Adobe PDF | Visualizar/Abrir |