Robust L-1 minimization object tracking via optical flow and object-centric method

Abstract

Object tracking has been studied in the field of computer vision and many approaches are proposed to solve challenging problems in the object tracking. Among many approaches, we are motivated by an existing tracking algorithm employing L-1 minimization method in particle filter framework. The L-1 minimization object tracking is one of state-of-the-art tracking algorithms, obtaining the robustness for illumination changes, occlusions and pose changes. However, the tracking algorithm often loses fast or complexly moving target object. In order to solve the problem, we suggest applying optical flow to the state model of particle filter framework. By adding calculated velocity information of the target object to the state model via optical flow, the performance of tracking for moving objects, especially fast or abruptly moving objects, is improved within our method. Moreover, we propose utilizing object-centric method in object tracking to improve the performance of the tracking algorithm and reduce the losses of target object during tracking. The object-centric method is inspired by object-centric spatial pooling that pools features from foreground and background separately. Object-centric method employs the difference of the Histogram of oriented gradients (HOG) features from the current frame and the first frame for both foreground and background features. We discuss the ineffectiveness of HOG features for object-centric method and suggest SIFT features instead of HOG features for the further studies.