WATCHING PEOPLE: ALGORITHMS TO STUDY HUMAN MOTION AND ACTIVITIES

Nowadays human motion analysis is one of the most active research topics in Computer Vision and it is receiving an increasing attention from both the industrial and scientific communities. The growing interest in human motion analysis is motivated by the increasing number of promising applications, ranging from surveillance, human–computer interaction, virtual reality to healthcare, sports, computer games and video conferencing, just to name a few. The aim of this thesis is to give an overview of the various tasks involved in visual motion analysis of the human body and to present the issues and possible solutions related to it. In this thesis, visual motion analysis is categorized into three major areas related to the interpretation of human motion: tracking of human motion using virtual pan-tilt-zoom (vPTZ) camera, recognition of human motions and human behaviors segmentation. In the field of human motion tracking, a virtual environment for PTZ cameras (vPTZ) is presented to overcame the mechanical limitations of PTZ cameras. The vPTZ is built on equirectangular images acquired by 360° cameras and it allows not only the development of pedestrian tracking algorithms but also the comparison of their performances. On the basis of this virtual environment, three novel pedestrian tracking algorithms for 360° cameras were developed, two of which adopt a tracking-by-detection approach while the last adopts a Bayesian approach. The action recognition problem is addressed by an algorithm that represents actions in terms of multinomial distributions of frequent sequential patterns of different length. Frequent sequential patterns are series of data descriptors that occur many times in the data. The proposed method learns a codebook of frequent sequential patterns by means of an apriori-like algorithm. An action is then represented with a Bag-of-Frequent-Sequential-Patterns approach. In the last part of this thesis a methodology to semi-automatically annotate behavioral data given a small set of manually annotated data is presented. The resulting methodology is not only effective in the semi-automated annotation task but can also be used in presence of abnormal behaviors, as demonstrated empirically by testing the system on data collected from children affected by neuro-developmental disorders.