Automatic human detection and tracking for robust video sequence annotation

Along this thesis, a novel and robust approach for automatic human annotation in long video sequences is addressed. This work defines a fully automatic pipeline that is able to deal with different types of sequences. The proposed system has been both designed and implemented following a divide and c...

Descripción completa

Detalles Bibliográficos
Autor: Llorca Queralt, Ramón
Tipo de recurso: tesis de maestría
Fecha de publicación:2014
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/87911
Acceso en línea:https://hdl.handle.net/2117/87911
Access Level:acceso embargado
Palabra clave:Video recording
Human face recognition (Computer science)
Face detection
face tracking
particle filters
local binary patterns
Viola & Jones
shot boundary detection and tracking-by-detection
automatic human annotation
Vídeo
Reconeixement facial (Informàtica)
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo
Descripción
Sumario:Along this thesis, a novel and robust approach for automatic human annotation in long video sequences is addressed. This work defines a fully automatic pipeline that is able to deal with different types of sequences. The proposed system has been both designed and implemented following a divide and conquer approach. First, a shot detector is used to divide the sequences in smaller ones. Then, humans are detected using a face detector based on the Viola & Jones algorithm. Once humans are detected, their faces are tracked using color-based particle filters and Local Binary Patterns (LBP). Several techniques and refinements have been implemented to improve the overall robustness of the system. Moreover, a track-by-detection technique is used to enhance the tracking accuracy. Finally, each human's track is annotated throughout every shot of the sequence. The performance of the global system is assessed in experiments with real sequences and compared against human made annotations. Furthermore, these annotated tracks set the groundwork for a future recognition system, that will complete the task of automatically annotating identities throughout sequences.