Tld and struck: A feature descriptors comparative study

Object tracking across multiple cameras is a very challenge issue in vision based monitoring applications. The selection of features is the first step to realize a reliable tracking algorithm.,In this work we analyse TLD and Struck, which are two of the most cited real-time visual trackers proposed in the literature in last years. They use two different feature extraction methodologies, Fern and Haar, respectively. The idea of this work is to compare performance of these well known visual tracking algorithms replacing their original feature characterization methods with local feature-based visual representations.,We test the improvement in terms of object detection and tracking performance grafting different features characterization into two completely different online tracker frameworks.,The used feature extraction methods are based on Local Binary Pattern (LBP), Local Gradient Pattern (LGP) and Histogram of Oriented Gradients (HOG). LGP is a novel detection methodology which is insensitive to global intensity variations like other representations such as local binary patterns (LBP).,The experimental results on well known benchmark sequences show as the feature extraction replacing improve the overall performances of the considered real-time visual trackers.