Real-time tracking based on weighted compressive tracking and a cognitive memory model

Abstract

Compressed tracking (CT) is a novel discriminative algorithm, treating the tracking problems as a binary classification between target and background. The method updates the appearance model, and locates the object in a compressed domain. However, CT algorithm can hardly deal with the condition that object suffers from serious occlusion or illumination variation. In this paper, we present a real-time tracking algorithm via combing weighted compressive tracking and a cognitive memory model. When the external environment is continuous and steady, the weighted compressive tracking (WCT) is employed due to its fast speed and high accuracy. Compared with CT treating each candidate window with the same prior probability, WCT imposes weighting factors on the basis of the Euclidean distance to the target window in previous frame. On the other hand, if the environment encounters a sudden change, a database is set to reserve different period tracking objects to locate the best match for the current frame, which is inspired by Atkinson-Shiffrin memory model (ASMM). 14 datasets in CVPR2013 Online Object Tracking Benchmark (OOTB) are used for extensive evaluation. The experimental results proved the superiority of the proposed algorithm with respect to other 5 state-of-the-art algorithms.