Multi-Object Tracking in Video Sequences

Object tracking in video sequences has found several different applications, such as surveillance, advanced interface, virtual reality, movement analysis, among others. However, the broadness of the spectrum of application is not caused by the existence of a general approach for object tracking. The inexistent generalization requires developers to build algorithms over several premises, in order to build robust solutions although in specific situations. To emphasize this issue, we present two approaches to object tracking, stating the premises of each method and the inoperability of one solution over the other context. In one hand, the dissertation presents an approach for vibrating line detection and tracking. Devices like accelerometers or load cells are normally used for measuring of vibrations in cables, wires or generic line structures. However, the installation of such equipment can be troublesome or even impossible in some cases, therefore new ways of capturing vibrating characteristics were developed. Computer vision allows a contactless technique to gather these measurements. However, the state-of-the-art algorithms either require a user initialization or a reduced description of the vibration. In this dissertation, a method for automatic line detection is presented, integrating a stable path approach, followed by an optical flow method for tracking the line‘s displacement. People tracking algorithms are an important subject within the object tracking field. These methods can usually be divided in several modules, consisting in initialization, tracking, pose estimation, among others. The tracking module matches people from one frame to another and therefore needs to make comparisons between two objects. This process often involves the use of an appearance model, i.e., a representation of the person, and requires from the developer a tradeoff decision between a reduced processing time and accurate characterization. An alternative method for people tracking is presented in this dissertation, consisting in the integration of an alternative appearance model which resorts to a local description using interest points, with the objective of an improvement in the description of a person.