Robust Visual Tracking Via Structural Local Dictionary Learning

Video tracking is a hot research problem in computer vision.It has shown great practical value in the field of intelligent unmanned aerial vehicle,video surveillance system,video compression,human-computer interaction and so on.In recent years,many good and robust video tracking algorithms have been proposed.However,designing a robust,accurate,and real-time video tracking algorithm is still a very challenging task in a variety of complex situations such as severe occlusion,clutter background,lighting changes,and shape changes.In this paper,we propose a novel video tracking algorithm based on structured local dictionary learning in the framework of particle filter.First,our algorithm uses an incremental updating algorithm to maintain an essentially low dimensional subspace on the target.The sub-space-based video tracking algorithm often achieves a better effect because the sub-space target object is less sensitive to local occlusion,illumination variation and other noise,and the low dimensional subspace we maintained is a less noisy and more robust subspace.On this basis,we construct a structured dictionary by using the local low rank features at multiple scales.Our structured dictionary effectively maintains the spatial layout and structural information of the target object,which makes our algorithm can more effectively deal with noise in different degrees of occlusion and shape changes.More importantly,we jointly consider all the candidate particles in the framework of particle filter.In this way,we remove some invalid local low-rank features due to the influence of noise,which suppress the random error caused by these local features,and make our tracking results more accurate.A large number of experiments show that our proposed algorithm has achieved very good experimental results in a variety of challenging data sets.