Human action identification in videos using descriptor with autonomous fragments and multilevel prediction

Author(s)

Marlon Fernandes de Alcantara

• Document type: Doctoral Thesis
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Computação
• Defense date: 2015-04-12

Examining board members

Hélio Pedrini; Alexandre Gonçalves da Silva; José Mário De Martino; Moacir Antonelli Ponti; Neucimar Jerônimo Leite

• Advisor: Hélio Pedrini

Abstract

Recent technological advances have provided devices with high processing power and storage capacities. Video cameras are found in several places, such as banks, airports, schools, supermarkets, streets, homes and industries. Despite this technological potential, most of the acquired videos are only stored and never analyzed. The flexibility in the use of cameras and computational tools allows their application in areas such as surveillance, strategic planning, crime prevention, manufacturing, traffic monitoring, among others. Video equipments have continuously improved, achieving high resolution rates and frames per second. However, most of the video analysis tasks are still performed by human operators, whose performance may be influenced by factors such stress and fatigue. In order to change such current scenario, this work proposes and evaluates the development of a methodology for identifying common human actions in videos by means of a CMSIP descriptor (Cumulative Motion Shape's Interest Points) applied to a multilevel prediction scheme with retraining. The approach is built by dividing the descriptor into portions that can be considered and interpreted independently by following distinct ways on the classification model, such that, in a later step, a central mechanism will be responsible for deciding which action is being observed in the video sequence. Our method has proved to be fast and with accuracy compatible to the state-of-the-art on known public data sets available in the literature, achieving 90% on Weizmann, KTH, MuHAVi and URADL data sets, whereas 82.6% on IXMAS data set. Furthermore, the developed prototype demonstrated to be a promising tool for real-time applications (AU)

• FAPESP's process: 12/20738-1 - Flexible-Context Activity Identification on Surveillance Cameras
• Grantee: Marlon Fernandes de Alcantara
• Support Opportunities: Scholarships in Brazil - Doctorate