PROJECT TITLE :

Transductive People Tracking in Unconstrained Surveillance

ABSTRACT:

Long-term tracking of individuals in unconstrained situations is still an open downside due to the absence of constant parts in the problem setting. The camera, when active, might move and the looks of each the background and the target may modification abruptly, resulting in the inadequacy of most standard tracking techniques. We tend to propose to use a learning approach that considers the tracking task as a semisupervised learning problem. Given few target samples, the aim is to go looking for the target occurrences within the video stream, reinterpreting the matter as label propagation on a similarity graph. We propose a solution primarily based on graph transduction that iteratively works frame by frame. Additionally, to avoid drifting, we tend to introduce an update strategy based mostly on an evolutionary clustering technique that chooses the visual templates that better describe target look, evolving the model throughout the processing of the video. Since we tend to model people's appearance by means that of covariance matrices on color and gradient info, our framework is directly related to structure learning on Riemannian manifolds. Tests on publicly obtainable knowledge sets and comparisons with state-of-the-art techniques enable us to conclude that our solution exhibits attention-grabbing performances in terms of tracking precision and recall in most of the thought of situations.