Multiple People Tracking and Gait Recognition for Video Surveillance

This thesis tackles the problems of multiple people tracking and gait recognition with a focus on occlusion handling in video surveillance. For the tracking, the proposed method adopts a hierarchical way of tracking along with a robust re-identification module, which produces affinity measures for data association. For gait recognition, two issues are addressed: 1) incomplete gait cycles, by presenting a fully convolutional network to restore complete gait cycles and 2) view variation, by proposing a variant of the non-negative matrix factorization algorithm.