Pedestrian detection in far infrared images

  1. Olmeda Reino, Daniel
Supervised by:
  1. Arturo de la Escalera Hueso Director
  2. José María Armingol Moreno Director

Defence university: Universidad Carlos III de Madrid

Fecha de defensa: 17 January 2014

  1. Andrés Iborra García Chair
  2. José Manuel Pastor García Secretary
  3. Joshué Pérez Rastelli Committee member

Type: Thesis


Detection of people in images is a relatively new field of research, but has been widely accepted. The applications are multiple, such as self-labeling of large databases, security systems and pedestrian detection in intelligent transportation systems. Within the latter, the purpose of a pedestrian detector from a moving vehicle is to detect the presence of people in the path of the vehicle. The ultimate goal is to avoid a collision between the two. This thesis is framed with the advanced driver assistance systems, passive safety systems that warn the driver of conditions that may be adverse. An advanced driving assistance system module, aimed to warn the driver about the presence of pedestrians, using computer vision in thermal images, is presented in this thesis. Such sensors are particularly useful under conditions of low illumination.The document is divided following the usual parts of a pedestrian detection system: development of descriptors that define the appearance of people in these kind of images, the application of these descriptors to full-sized images and temporal tracking of pedestrians found. As part of the work developed in this thesis, database of pedestrians in the far infrared spectrum is presented. This database has been used in developing an evaluation of pedestrian detection systems as well as for the development of new descriptors. These descriptors use techniques for the systematic description of the shape of the pedestrian as well as methods to achieve invariance to contrast, illumination or ambient temperature. The descriptors are analyzed and modified to improve their performance in a detection problem, where potential candidates are searched for in full size images. Finally, a method for tracking the detected pedestrians is proposed to reduce the number of miss-detections that occurred at earlier stages of the algorithm. --