Research Of Appearance Model Based Object Tracking

Visual object tracking is one of the fundamental problems in computer vision,which facilitates various higher-level applications object recognition,surveillance and event analysis.According to the object state at the initial frame from video,object tracking should estimate the object optimal center locations and the bounding boxes at the remainder frames.Although object tracking has been studied for several years,it is still challenging to design a robust and effective algorithm due to various critical situations,such as complex motions,illumination variations,partial occlusions,and background clutters,etc.For robust tracking under these critical situations,it is essential to built effective appearance models.Effective appearance model is crucial and difficult for a robust object tracking.Many appearance models have been designed to represent the target for object tracking,and these tracking methods can be roughly classified into two categories according to their appearance models: bounding box models and patch models.A bounding box model utilizes the entire bounding box of the target to extract the features,which are more discriminative for trackers separating the target form cluttered background coarsely.On the other hand,a patch model divides the bounding box into multiple smaller patches and extracts features for each patch separately.Although small patches can provide many spatial details of the target,they may be no clearly distinguishable from background patches.So object tracking algorithms often take the background as the target,which makes the objects shifting and the tracking failing.In order to overcome the aforementioned problems,we propose two robust trackers according to the appearance model.The first one is a robust tracker based on the diverse multiple templates,and the other one is a novel structural multi-size patch appearance model to achieve reliable and accurate tracking.With regard to the tracker based on the diverse multiple templates,the key idea is to maintain the diversity of multiple templates,which makes the templates in appearance model different and crucial.And the multiple templates can provide more apriori information of the target to make the tracking results more robust and accurate in the process of tacking.With regard to the tracker based on the structural multi-size patch appearance model,its key idea to build the relationships among different size patches for the fusion of bounding box models and patch models.In the proposed appearance model,the large patches represent the bounding box model and the small patches express the patch model.According to the multi-size patch correspondences,a coarse-to-fine fusion strategy is employed to estimate the accurate position and scale of the target.We evaluate the proposed trackers on several publicly available challenging video sequences from dataset CVPR2013,OTB100 and VOT2014,and we compare our tracker with the state-of-the-art tracking algorithms.Experimental results on large set of video sequences showed that the proposed methods outperforms state-of-the-art trackers.