Visual Analysis of Pedestrian Motion

We describe progress towards visual analysis of pedestrian motion. While trajectories of humans on foot are stochastic in nature, in a constrianed situation underlying patterns of motion can be identified. The work presented in this report focuses on movement of people through scenes which are under visual surveillance. In this case, analysis of continuous video footage provides a history of many observed trajectories which, to a human observer, readily reveals common properties of paths taken. We study the problems associated with automatically recognising such patterns, and also utilising this information to aid other automated tasks. We begin with a review of previous attempts to model pedestrian activity in visual surveillance, highlighting strengths and weaknesses of current approaches. We also review algorithms which are key components of a visual system for tracking people from mobile cameras. The state of the art in pedestrian detection is reviewed, and details of our implementation of an algorithm designed for close to real time detection are discussed. The main part of this report describes our novel representation of pedestrian motion patterns. The major strength of this method lies in our use of Gaussian processes for regression which allows us to be explicit about the uncertainty in pedestrian motion. The approach is nonparametric and so full use can be made of the vast quantities of data which surveillance footage provides. We exemplify the use of this model for long term prediction of target motion, and illustrate how the learned model produces more accurate predictions than a typical memoryless system. Finally, we propose a plan for future work comprising both short term goals and long term research direction. In the short term, plans to extend the scene activity model are presented along with applications in visual surveillance. In the long term, we propose to extend the analysis of human motion to less constrained situations where a mobile camera is used.