Multiple object tracking in real-time

Abstract

Much research has been undertaken in the broad field of Multiple Object Tracking. Systems developed to date have encountered success in tracking almost every variety of moving object. The majority of the research in the field is aimed towards producing accurate tracking models for peoples' faces and bodies, for access authentification and for security type applications. This thesis proposes two real-time systems to cater for both the tracking of single objects of any size and description and the typical application of tracking multiple faces in a dynamic scene. The techniques discussed in this thesis include a comparison of three differing methods of motion localisation using Temporal filtering, skin colour detection and background subtraction. The techniques of template matching using Relative Projection Histograms, and motion prediction using Kalman filtering are trialed in a real-time tracking environment with excellent tracking results at slightly constrained scales. The application of a comprehensive neural network face detector is also discussed and developed. This face detector is typical of the type used in most recent papers concerned with tracking multiple faces in dynamic scenes. Results and conclusions on the effectiveness of the three motion localisation techniques are presented, along with an analysis of the results of template matching used for object tracking. A partially completed face detector is also presented as a direction for future work in the field of Multiple Object Tracking.