Research On Superpixel Tracking With The Integration Of Spatio-Temporal Context

Visual Tracking is one of the most important research tasks in computer vision. Forthe past few years, target object tracking attracts lots of attentions of researchers andenterprise all over the world due to its numerous applications, such as surveillance,intelligent traffic and human computer interaction. Object tracking in computer vision hasa great potential for development.In this paper，we first introduce the research background and significance for theobject tracking. And then we describe the Research Status of this field, and select andsummarize the classical algorithm and the key technology for object tracking. At the sametime, we indicate the current focuses and the challenging factors in the field of visualtracking. Next, a novel superpixel tracking algorithm is proposed which adopt themid-level visual features for tracking. As far as I can see, the superpixel, which is therepresentation of mid-level visual features, is seldom used in object tracking field before.The spatio-temporal context information is also used in order to achieve more accuratetracking results. Finally, a lot of experiments demonstrate that the proposed superpixeltracking algorithm has the ability of to handle multiple effect factors, e.g. illumination andscale variation, heavy occasion and rotation, simultaneously, and more robust andeffective than previous classic tracking algorithms.The research on visual tracking technology has been developed several decades. Aseries of outstanding methods have been proposed, but most of them perform well onsequences which suffer from only one or two effect factors. However, all the effect factorsoccur simultaneously in reality scenes. It remains a challenging problem. The key is toconstruct a quite effective appearance model, which is adaptive to all changes caused byevery effect factor. Most tracking methods adopt the high-level visual features (e.g. Haar-like) or the raw low-level visual features. Unlike them, our superpixel trackingalgorithm is based on the mid-level visual features, which takes advantage of super pixel.Therefore, the proposed algorithm is more adaptive and robust. In this paper, objecttracking is related to computing a location confidence map, and the target location isdetermined to the place where the confidence is maximized.In the last part, proposed superpixel tracking algorithm is experimented on fivebenchmark databases and compared with three outstanding tracking algorithms.Furthermore, the performance of the proposed superpixel tracking algorithm is evaluatedfrom qualitative and quantitative aspects, respectively.